# =============================================================
# File: backend/api/services/agent_orchestrator.py
# =============================================================
"""
Agent Orchestrator (integrated with enterprise RedFlagDetector)
Place at: backend/api/services/agent_orchestrator.py
"""
from __future__ import annotations
import asyncio
import json
import os
import re
from typing import List, Dict, Any, Optional
import logging
from ..models.agent import AgentRequest, AgentDecision, AgentResponse
from ..models.redflag import RedFlagMatch
from .redflag_detector import RedFlagDetector
from .intent_classifier import IntentClassifier
from .tool_selector import ToolSelector
from .llm_client import LLMClient
from ..mcp_clients.mcp_client import MCPClient
from .tool_scoring import ToolScoringService
from ..storage.analytics_store import AnalyticsStore
from .result_merger import merge_parallel_results, format_merged_context_for_prompt
from .tool_metadata import validate_tool_output, get_tool_schema
from .query_cache import get_cache
from .query_expander import QueryExpander
from .context_engineer import ContextEngineer
import time
logger = logging.getLogger(__name__)
from dotenv import load_dotenv
load_dotenv()
class AgentOrchestrator:
def __init__(self, rag_mcp_url: str, web_mcp_url: str, admin_mcp_url: str):
self.mcp = MCPClient(rag_mcp_url, web_mcp_url, admin_mcp_url)
# Groq-only LLM client
self.llm = LLMClient(api_key=os.getenv("GROQ_API_KEY"), model=os.getenv("GROQ_MODEL"))
# pass admin_mcp_url so detector can call back
self.redflag = RedFlagDetector(
supabase_url=os.getenv("SUPABASE_URL"),
supabase_key=os.getenv("SUPABASE_SERVICE_KEY"),
admin_mcp_url=admin_mcp_url
)
self.intent = IntentClassifier(llm_client=self.llm)
self.selector = ToolSelector(llm_client=self.llm)
self.tool_scorer = ToolScoringService()
self.query_expander = QueryExpander(llm_client=self.llm)
self.cache = get_cache()
self.context_engineer = ContextEngineer(llm_client=self.llm)
self._analytics: Optional[AnalyticsStore] = None
self._analytics_disabled = os.getenv("ANALYTICS_DISABLED", "").lower() in {"1", "true", "yes"}
self._analytics_failed = False
self._log_analytics_backend_once()
def _log_analytics_backend_once(self) -> None:
if getattr(AgentOrchestrator, "_analytics_backend_logged", False):
return
if self._analytics_disabled:
logger.info("Analytics: Disabled via ANALYTICS_DISABLED")
else:
store = self._get_analytics()
if store is None:
# Only log if credentials might be missing (not if package is missing)
import os
if os.getenv("SUPABASE_URL") and os.getenv("SUPABASE_SERVICE_KEY"):
logger.warning("Analytics: Disabled (Supabase initialization failed)")
else:
logger.debug("Analytics: Disabled (Supabase not configured)")
elif store.use_supabase:
logger.info("Analytics: Using Supabase backend")
else:
logger.warning("Analytics: Using fallback backend")
AgentOrchestrator._analytics_backend_logged = True
def _get_analytics(self) -> Optional[AnalyticsStore]:
if self._analytics_disabled or self._analytics_failed:
return None
if self._analytics is not None:
return self._analytics
try:
self._analytics = AnalyticsStore()
except RuntimeError as exc:
# Only log at warning level if credentials are configured (actual error)
# Otherwise log at debug level (expected when Supabase is not configured)
import os
if os.getenv("SUPABASE_URL") and os.getenv("SUPABASE_SERVICE_KEY"):
logger.warning("Analytics disabled: %s", str(exc).split('\n')[0]) # Only first line
else:
logger.debug("Analytics disabled: %s", str(exc).split('\n')[0])
self._analytics_failed = True
self._analytics = None
except Exception as exc: # pragma: no cover - unexpected initialization failures
logger.debug("Analytics unexpected init failure: %s", exc)
self._analytics_failed = True
self._analytics = None
return self._analytics
def _analytics_log_tool_usage(self, **kwargs: Any) -> None:
analytics = self._get_analytics()
if not analytics:
return
try:
analytics.log_tool_usage(**kwargs)
except Exception as exc: # pragma: no cover - analytics failures should not break flow
logger.debug("AgentOrchestrator tool analytics failed: %s", exc)
def _analytics_log_agent_query(self, **kwargs: Any) -> None:
analytics = self._get_analytics()
if not analytics:
return
try:
analytics.log_agent_query(**kwargs)
except Exception as exc: # pragma: no cover
logger.debug("AgentOrchestrator agent query analytics failed: %s", exc)
def _analytics_log_rag_search(self, **kwargs: Any) -> None:
analytics = self._get_analytics()
if not analytics:
return
try:
analytics.log_rag_search(**kwargs)
except Exception as exc: # pragma: no cover
logger.debug("AgentOrchestrator RAG analytics failed: %s", exc)
def _analytics_log_redflag_violation(self, **kwargs: Any) -> None:
analytics = self._get_analytics()
if not analytics:
return
try:
analytics.log_redflag_violation(**kwargs)
except Exception as exc: # pragma: no cover
logger.debug("AgentOrchestrator redflag analytics failed: %s", exc)
def _cache_response(self, req: AgentRequest, response: AgentResponse, skip_cache: bool = False):
"""Cache a response if appropriate."""
if skip_cache or req.message.startswith("admin:") or len(req.message) < 3:
return
try:
self.cache.set(req.message, req.tenant_id, {
"text": response.text,
"decision": response.decision.dict() if response.decision else None,
"tool_traces": response.tool_traces,
"reasoning_trace": response.reasoning_trace
})
except Exception as e:
logger.debug(f"Failed to cache response: {e}")
async def handle(self, req: AgentRequest) -> AgentResponse:
start_time = time.time()
reasoning_trace: List[Dict[str, Any]] = []
reasoning_trace.append({
"step": "request_received",
"tenant_id": req.tenant_id,
"user_id": req.user_id,
"message_preview": req.message[:120]
})
# Context Engineering: Write to scratchpad
self.context_engineer.write_to_scratchpad(
f"User query: {req.message[:200]}",
category="user_query"
)
# Check cache first (skip for admin queries and rule checks)
cached_response = self.cache.get(req.message, req.tenant_id)
if cached_response:
reasoning_trace.append({
"step": "cache_hit",
"cached": True
})
return AgentResponse(
text=cached_response.get("text", ""),
decision=cached_response.get("decision"),
tool_traces=cached_response.get("tool_traces", []),
reasoning_trace=reasoning_trace + cached_response.get("reasoning_trace", [])
)
# 1) FIRST: Check admin rules - if any rule matches, respond according to rule
matches: List[RedFlagMatch] = await self.redflag.check(req.tenant_id, req.message)
reasoning_trace.append({
"step": "admin_rules_check",
"match_count": len(matches),
"matches": [m.__dict__ for m in matches]
})
if matches:
# Log all rule matches
for match in matches:
self._analytics_log_redflag_violation(
tenant_id=req.tenant_id,
rule_id=match.rule_id,
rule_pattern=match.pattern,
severity=match.severity,
matched_text=match.matched_text,
confidence=match.confidence,
message_preview=req.message[:200],
user_id=req.user_id
)
# Categorize rules: brief response rules vs blocking rules
brief_response_rules = []
blocking_rules = []
for match in matches:
rule_text = (match.description or match.pattern or "").lower()
is_brief_rule = (
match.severity == "low" and (
"greeting" in rule_text or
"brief" in rule_text or
"simple response" in rule_text or
"keep.*response.*brief" in rule_text or
"do not.*verbose" in rule_text or
"respond.*briefly" in rule_text
)
)
if is_brief_rule:
brief_response_rules.append(match)
else:
blocking_rules.append(match)
# Handle brief response rules (greetings, etc.) - return immediately
if brief_response_rules and not blocking_rules:
# Return brief response without proceeding to normal flow
brief_responses = [
"Hello! How can I help you today?",
"Hi there! What can I assist you with?",
"Hello! I'm here to help. What do you need?",
"Hi! How can I assist you?"
]
import random
brief_response = random.choice(brief_responses)
reasoning_trace.append({
"step": "brief_response_rule_matched",
"action": "brief_response",
"matched_rules": [m.rule_id for m in brief_response_rules],
"message": "Brief response rule matched, returning brief response (skipping normal flow)"
})
total_latency_ms = int((time.time() - start_time) * 1000)
self._analytics_log_agent_query(
tenant_id=req.tenant_id,
message_preview=req.message[:200],
intent="greeting",
tools_used=[],
total_tokens=len(brief_response) // 4,
total_latency_ms=total_latency_ms,
success=True,
user_id=req.user_id
)
return AgentResponse(
text=brief_response,
decision=AgentDecision(action="respond", tool=None, tool_input=None, reason="brief_response_rule"),
reasoning_trace=reasoning_trace
)
# Handle blocking rules (security, compliance, etc.) - block and return immediately
if blocking_rules:
# Notify admin asynchronously
try:
await self.redflag.notify_admin(req.tenant_id, blocking_rules, source_payload={"message": req.message, "user_id": req.user_id})
except Exception:
pass
decision = AgentDecision(
action="block",
tool="admin",
tool_input={"violations": [m.__dict__ for m in blocking_rules]},
reason="admin_rule_violation"
)
# Build detailed prompt for LLM to generate natural red flag response
violations_details = []
for i, m in enumerate(blocking_rules, 1):
rule_name = m.description or m.pattern or "Policy violation"
detail = f"{i}. **{rule_name}** (Severity: {m.severity})"
if m.matched_text:
detail += f"\n - Detected phrase: \"{m.matched_text}\""
violations_details.append(detail)
llm_prompt = f"""A user made the following request: "{req.message}"
However, this request violates company policies. The following policy violations were detected:
{chr(10).join(violations_details)}
Your task: Write a clear, professional, and empathetic response to inform the user that:
1. Their request cannot be processed due to policy violations
2. Which specific policy was violated (mention it naturally)
3. The incident has been logged for security review
4. They should contact an administrator if they need assistance or believe this is an error
Write a natural, conversational response (2-4 sentences) that feels helpful rather than robotic. Be professional but understanding.
Response:"""
# Generate LLM response for red flag
try:
llm_response = await self.llm.simple_call(llm_prompt, temperature=min(req.temperature + 0.2, 0.7)) # Slightly higher temp for more natural response
llm_response = llm_response.strip()
# Add warning emoji if not present
if not llm_response.startswith("⚠️") and not llm_response.startswith("🚨"):
llm_response = f"⚠️ {llm_response}"
except Exception as e:
# Fallback to a simple message if LLM fails
summary = "; ".join(
f"{m.description or m.pattern}"
for m in matches
)
llm_response = f"⚠️ I'm unable to process your request because it violates our company policy: {summary}. This incident has been logged. Please contact your administrator if you need assistance."
total_latency_ms = int((time.time() - start_time) * 1000)
# Log LLM usage for red flag response
estimated_tokens = len(llm_response) // 4 + len(llm_prompt) // 4
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="llm",
latency_ms=total_latency_ms,
tokens_used=estimated_tokens,
success=True,
user_id=req.user_id
)
self._analytics_log_agent_query(
tenant_id=req.tenant_id,
message_preview=req.message[:200],
intent="admin",
tools_used=["admin", "llm"],
total_tokens=estimated_tokens,
total_latency_ms=total_latency_ms,
success=False,
user_id=req.user_id
)
response = AgentResponse(
text=llm_response,
decision=decision,
tool_traces=[{"redflags": [m.__dict__ for m in blocking_rules]}],
reasoning_trace=reasoning_trace
)
# Don't cache admin rule violations
return response
# 2) ONLY IF NO RULES MATCHED: Proceed with normal flow (intent classification, RAG, etc.)
# 2.1) Optional: Try to rewrite message if it might violate rules (preventive self-correction)
# Note: This is a lighter check - we already blocked above if rules matched
# This is for edge cases where we want to proactively improve the message
safe_message = req.message # Default to original
intent = await self.intent.classify(req.message)
reasoning_trace.append({
"step": "intent_detection",
"intent": intent
})
# 2.5) Pre-fetch RAG results if available (for tool selector context)
# BUT: Skip RAG pre-fetch for news/current events queries (they need web search, not RAG)
rag_prefetch = None
rag_results = []
# Detect news queries early to skip RAG pre-fetch
# Make detection more aggressive - check for "news" keyword first
msg_lower = req.message.lower().strip()
# Primary detection: if "news" is in the message, it's almost certainly a news query
has_news_keyword = "news" in msg_lower
# Exclude common non-news phrases that contain "news" but aren't news queries
non_news_phrases = [
"what is", "what's", "explain", "tell me about", "define",
"how does", "how do", "what are", "what does", "what can"
]
is_general_question = any(phrase in msg_lower for phrase in non_news_phrases)
freshness_keywords = ["latest", "today", "current", "recent",
"now", "updates", "breaking", "trending", "happening",
"what's new", "what is new", "what happened"]
news_patterns = [
r"latest news", r"current news", r"today's news", r"breaking news",
r"news about", r"news on", r"news of", r"what's happening",
r"what happened", r"recent news", r"news update"
]
# If "news" keyword is present AND it's not a general question, it's a news query
# Otherwise check for other freshness indicators
is_news_query = (has_news_keyword and not is_general_question) or \
(any(k in msg_lower for k in freshness_keywords) and not is_general_question) or \
any(re.search(p, msg_lower) for p in news_patterns)
# LLM-based detection for edge cases (if keyword-based detection is uncertain)
# Only use LLM if it's a short query and we're uncertain
if not is_news_query and len(msg_lower.split()) <= 5 and not is_general_question:
# For short queries, use LLM to check if it's a news query
try:
llm_check_prompt = f"""Is the following query asking for current news or recent events? Answer only "yes" or "no".
Query: "{req.message}"
Answer:"""
llm_response = await self.llm.simple_call(llm_check_prompt, temperature=0.0)
if "yes" in llm_response.lower():
is_news_query = True
reasoning_trace.append({
"step": "news_query_detection_llm",
"detected": True,
"llm_confirmed": True
})
except Exception as e:
logger.debug(f"LLM news detection failed: {e}")
# Log detection for debugging
if is_news_query:
reasoning_trace.append({
"step": "news_query_detection",
"detected": True,
"message": req.message,
"has_news_keyword": has_news_keyword,
"matched_keywords": [k for k in freshness_keywords if k in msg_lower]
})
# Only pre-fetch RAG if it's NOT a news query
if not is_news_query:
try:
# Try to pre-fetch RAG to help tool selector make better decisions
rag_start = time.time()
rag_prefetch = await self.mcp.call_rag(req.tenant_id, req.message)
rag_latency_ms = int((time.time() - rag_start) * 1000)
if isinstance(rag_prefetch, dict):
rag_results = rag_prefetch.get("results") or rag_prefetch.get("hits") or []
# Log RAG search event
hits_count = len(rag_results)
avg_score = None
top_score = None
if rag_results:
scores = [h.get("score", 0.0) for h in rag_results if isinstance(h, dict) and "score" in h]
if scores:
avg_score = sum(scores) / len(scores)
top_score = max(scores)
self._analytics_log_rag_search(
tenant_id=req.tenant_id,
query=req.message[:500],
hits_count=hits_count,
avg_score=avg_score,
top_score=top_score,
latency_ms=rag_latency_ms
)
# Log tool usage
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="rag",
latency_ms=rag_latency_ms,
success=True,
user_id=req.user_id
)
reasoning_trace.append({
"step": "rag_prefetch",
"status": "ok",
"hit_count": len(rag_results),
"latency_ms": rag_latency_ms
})
except Exception as pref_err:
# If RAG fails, continue without it
rag_latency_ms = 0 # 0 for failed
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="rag",
latency_ms=rag_latency_ms,
success=False,
error_message=str(pref_err)[:200],
user_id=req.user_id
)
reasoning_trace.append({
"step": "rag_prefetch",
"status": "error",
"error": str(pref_err)
})
rag_prefetch = None
else:
# News query detected - skip RAG pre-fetch
reasoning_trace.append({
"step": "rag_prefetch",
"status": "skipped",
"reason": "news_query_detected"
})
tool_scores = self.tool_scorer.score(req.message, intent, rag_results)
reasoning_trace.append({
"step": "tool_scoring",
"scores": tool_scores
})
# 3) Tool selection (hybrid) - pass RAG results, memory, and admin violations in context
# Context Engineering: Compress conversation history if too long (Anthropic's compaction)
# Use tool result clearing first (safest), then full compaction if needed
if req.conversation_history and len(req.conversation_history) > 10:
# Check token usage
total_chars = sum(len(str(m.get("content", ""))) for m in req.conversation_history)
estimated_tokens = total_chars // 4
# Compress if approaching context limit (80% threshold)
if estimated_tokens > 8000: # ~80% of typical 10k context
compressed_history = await self.context_engineer.compress_if_needed(
req.conversation_history,
max_tokens=6000, # Target 60% after compression
use_compaction=True
)
req.conversation_history = compressed_history
reasoning_trace.append({
"step": "context_compaction",
"original_length": len(req.conversation_history),
"compressed_length": len(compressed_history),
"compressed": len(compressed_history) < len(req.conversation_history),
"strategy": "anthropic_compaction"
})
# Get recent memory for context-aware routing
from backend.mcp_server.common.memory import get_recent
session_id = req.conversation_history[-1].get("session_id") if req.conversation_history else None
recent_memory = []
if session_id:
recent_memory = get_recent(session_id)
# Context Engineering: Select relevant memories
if recent_memory:
selected_memories = await self.context_engineer.select_context(
req.message,
{"memories": recent_memory}
)
recent_memory = selected_memories.get("memories", recent_memory)
# Get admin violations if any
admin_violations = []
if hasattr(self, 'redflag') and self.redflag:
# Check if there were any violations detected
# (This would be set during redflag checking earlier in the flow)
pass # Admin violations are checked separately
# FORCE web search for news queries - bypass tool selector entirely
# Also ensure rag_results is empty for news queries (double-check)
if is_news_query:
rag_results = [] # Force empty - no RAG results for news queries
from ..models.agent import AgentDecision
# Enhance query for better web search results
web_query = req.message
# Handle ambiguous short queries like "latest news about Al" or "atest news about Al"
# Try to expand with common interpretations
query_words = web_query.lower().split()
if len(query_words) <= 4:
# Extract the topic (word after "about" or last word)
topic = None
if "about" in query_words:
about_idx = query_words.index("about")
if about_idx + 1 < len(query_words):
topic = query_words[about_idx + 1]
elif len(query_words) >= 2:
# Last word might be the topic
topic = query_words[-1]
# If topic is very short (1-2 letters), it's likely ambiguous - expand it
if topic and len(topic) <= 2:
# Common expansions for "Al"
if topic == "al":
# Try multiple interpretations
web_query = f"{' '.join(query_words[:-1])} artificial intelligence AI"
elif topic == "ai":
web_query = f"{' '.join(query_words[:-1])} artificial intelligence"
# If still short, add "news" keyword if missing
if "news" not in web_query.lower() and len(web_query.split()) <= 3:
web_query = f"{web_query} news latest"
decision = AgentDecision(
action="call_tool",
tool="web",
tool_input={"query": web_query},
reason=f"news_query_forced_web_search (original: {req.message})"
)
reasoning_trace.append({
"step": "tool_selection",
"decision": decision.dict(),
"note": "news_query_bypassed_selector_forced_web",
"rag_results_forced_empty": True,
"web_query": web_query
})
else:
ctx = {
"tenant_id": req.tenant_id,
"rag_results": rag_results,
"tool_scores": tool_scores,
"memory": recent_memory, # Context-aware routing: recent tool outputs
"admin_violations": admin_violations # Context-aware routing: admin rule severity
}
decision = await self.selector.select(intent, req.message, ctx)
reasoning_trace.append({
"step": "tool_selection",
"decision": decision.dict(),
"context_scores": tool_scores
})
tool_traces: List[Dict[str, Any]] = []
# 4) Handle multi-step tool execution
if decision.action == "multi_step" and decision.tool_input:
steps = decision.tool_input.get("steps", [])
if steps:
return await self._execute_multi_step(
req,
steps,
decision,
tool_traces,
reasoning_trace,
rag_prefetch
)
# 5) Execute single tool
tools_used = []
total_tokens = 0
if decision.action == "call_tool" and decision.tool:
try:
if decision.tool == "rag":
# Use autonomous retry with self-correction
rag_query = decision.tool_input.get("query") if decision.tool_input else req.message
rag_start = time.time()
rag_resp = await self.rag_with_repair(
query=rag_query,
tenant_id=req.tenant_id,
original_threshold=0.3,
reasoning_trace=reasoning_trace,
user_id=req.user_id
)
rag_latency_ms = int((time.time() - rag_start) * 1000)
tools_used.append("rag")
# Validate and format RAG output to conform to schema
rag_formatted = self._format_tool_output("rag", rag_resp, rag_latency_ms)
# Context Engineering: Compress tool output if needed
rag_formatted = await self.context_engineer.compressor.compress_tool_output("rag", rag_formatted)
tool_traces.append({"tool": "rag", "response": rag_formatted})
hits = self._extract_hits(rag_formatted)
# Calculate scores for logging
hits_count = len(hits)
avg_score = rag_formatted.get("avg_score")
top_score = rag_formatted.get("top_score")
reasoning_trace.append({
"step": "tool_execution",
"tool": "rag",
"hit_count": hits_count,
"top_score": top_score,
"avg_score": avg_score,
"summary": self._summarize_hits(rag_formatted, limit=2)
})
prompt = self._build_prompt_with_rag(req, rag_formatted)
llm_start = time.time()
llm_out = await self.llm.simple_call(prompt, temperature=req.temperature)
llm_latency_ms = int((time.time() - llm_start) * 1000)
tools_used.append("llm")
# Estimate tokens (rough: ~4 chars per token)
estimated_tokens = len(llm_out) // 4 + len(prompt) // 4
total_tokens += estimated_tokens
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="llm",
latency_ms=llm_latency_ms,
tokens_used=estimated_tokens,
success=True,
user_id=req.user_id
)
reasoning_trace.append({
"step": "llm_response",
"mode": "rag_synthesis",
"latency_ms": llm_latency_ms,
"estimated_tokens": estimated_tokens
})
total_latency_ms = int((time.time() - start_time) * 1000)
self._analytics_log_agent_query(
tenant_id=req.tenant_id,
message_preview=req.message[:200],
intent=intent,
tools_used=tools_used,
total_tokens=total_tokens,
total_latency_ms=total_latency_ms,
success=True,
user_id=req.user_id
)
return AgentResponse(text=llm_out, decision=decision, tool_traces=tool_traces, reasoning_trace=reasoning_trace)
if decision.tool == "web":
# CRITICAL: For news queries, ensure RAG results are NEVER used
msg_check_web = req.message.lower()
is_news_web = "news" in msg_check_web or any(k in msg_check_web for k in ["latest", "breaking", "current", "recent", "today"])
if is_news_web:
# Force clear any RAG context - news queries should NEVER use RAG
rag_results = []
reasoning_trace.append({
"step": "web_tool_execution",
"note": "news_query_confirmed_rag_results_cleared_before_web_search"
})
# Use autonomous retry with query rewriting
web_query = decision.tool_input.get("query") if decision.tool_input else req.message
web_start = time.time()
web_resp = await self.web_with_repair(
query=web_query,
tenant_id=req.tenant_id,
reasoning_trace=reasoning_trace,
user_id=req.user_id
)
web_latency_ms = int((time.time() - web_start) * 1000)
tools_used.append("web")
# Validate and format Web output to conform to schema
web_formatted = self._format_tool_output("web", web_resp, web_latency_ms)
# Context Engineering: Compress tool output if needed
web_formatted = await self.context_engineer.compressor.compress_tool_output("web", web_formatted)
tool_traces.append({"tool": "web", "response": web_formatted})
hits_count = len(self._extract_hits(web_formatted))
reasoning_trace.append({
"step": "tool_execution",
"tool": "web",
"hit_count": hits_count,
"summary": self._summarize_hits(web_formatted, limit=2),
"is_news_query": is_news_web
})
# ALWAYS use web prompt builder for web search results
# Never use RAG prompt builder, even if web results are empty
if hits_count == 0 and is_news_web:
# Empty web results for news query - provide helpful guidance
prompt = (
f"You are an assistant helping tenant {req.tenant_id}.\n\n"
f"## User Question\n{req.message}\n\n"
f"## Context\n"
f"I searched for the latest news about this topic, but didn't find specific recent results in my web search.\n\n"
f"## Your Task\n"
f"Provide helpful information about what the user might be looking for. "
f"If you have general knowledge about the topic, share it. "
f"Be honest that I don't have access to the very latest breaking news right now, but provide what context you can. "
f"Suggest that the user try:\n"
f"- Checking major news websites directly (BBC, CNN, Reuters, etc.)\n"
f"- Trying a more specific search query\n"
f"- Using a news aggregator service\n\n"
f"IMPORTANT: Do NOT say 'There is no mention of X in the provided context' - instead provide helpful general information or suggest where to find current news.\n\n"
f"Provide a helpful response now:"
)
else:
# Use web prompt builder (never RAG)
prompt = self._build_prompt_with_web(req, web_formatted)
llm_start = time.time()
llm_out = await self.llm.simple_call(prompt, temperature=req.temperature)
llm_latency_ms = int((time.time() - llm_start) * 1000)
tools_used.append("llm")
estimated_tokens = len(llm_out) // 4 + len(prompt) // 4
total_tokens += estimated_tokens
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="llm",
latency_ms=llm_latency_ms,
tokens_used=estimated_tokens,
success=True,
user_id=req.user_id
)
reasoning_trace.append({
"step": "llm_response",
"mode": "web_synthesis",
"latency_ms": llm_latency_ms,
"estimated_tokens": estimated_tokens
})
total_latency_ms = int((time.time() - start_time) * 1000)
self._analytics_log_agent_query(
tenant_id=req.tenant_id,
message_preview=req.message[:200],
intent=intent,
tools_used=tools_used,
total_tokens=total_tokens,
total_latency_ms=total_latency_ms,
success=True,
user_id=req.user_id
)
return AgentResponse(text=llm_out, decision=decision, tool_traces=tool_traces, reasoning_trace=reasoning_trace)
if decision.tool == "admin":
admin_start = time.time()
admin_resp = await self.mcp.call_admin(req.tenant_id, decision.tool_input.get("query") if decision.tool_input else req.message)
admin_latency_ms = int((time.time() - admin_start) * 1000)
tools_used.append("admin")
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="admin",
latency_ms=admin_latency_ms,
success=True,
user_id=req.user_id
)
# Validate and format Admin output to conform to schema
admin_formatted = self._format_tool_output("admin", admin_resp, admin_latency_ms)
tool_traces.append({"tool": "admin", "response": admin_formatted})
reasoning_trace.append({
"step": "tool_execution",
"tool": "admin",
"status": "completed",
"latency_ms": admin_latency_ms
})
total_latency_ms = int((time.time() - start_time) * 1000)
self._analytics_log_agent_query(
tenant_id=req.tenant_id,
message_preview=req.message[:200],
intent=intent,
tools_used=tools_used,
total_tokens=0,
total_latency_ms=total_latency_ms,
success=True,
user_id=req.user_id
)
return AgentResponse(text=json.dumps(admin_resp), decision=decision, tool_traces=tool_traces, reasoning_trace=reasoning_trace)
if decision.tool == "llm":
# Check if this is a news query - if so, force web search instead
msg_lower_llm = req.message.lower()
freshness_keywords_llm = ["latest", "today", "news", "current", "recent",
"now", "updates", "breaking", "trending", "happening"]
news_patterns_llm = [
r"latest news", r"current news", r"today's news", r"breaking news",
r"news about", r"news on", r"news of"
]
is_news_query_llm = any(k in msg_lower_llm for k in freshness_keywords_llm) or \
any(re.search(p, msg_lower_llm) for p in news_patterns_llm)
# Force web search for news queries even if tool selector chose "llm"
if is_news_query_llm:
try:
web_query = req.message
if len(web_query.split()) <= 4:
if "news" not in msg_lower_llm:
web_query = f"{web_query} news latest"
web_start = time.time()
web_resp = await self.web_with_repair(
query=web_query,
tenant_id=req.tenant_id,
reasoning_trace=reasoning_trace,
user_id=req.user_id
)
web_latency_ms = int((time.time() - web_start) * 1000)
tools_used.append("web")
web_formatted = self._format_tool_output("web", web_resp, web_latency_ms)
# Context Engineering: Compress tool output if needed
web_formatted = await self.context_engineer.compressor.compress_tool_output("web", web_formatted)
tool_traces.append({"tool": "web", "response": web_formatted})
hits_count = len(self._extract_hits(web_formatted))
reasoning_trace.append({
"step": "tool_execution",
"tool": "web",
"hit_count": hits_count,
"note": "forced_web_for_news_in_llm_path"
})
if hits_count == 0:
prompt_for_llm = (
f"You are an assistant helping tenant {req.tenant_id}.\n\n"
f"## User Question\n{req.message}\n\n"
f"## Context\n"
f"I attempted to search for the latest news about this topic, but didn't find specific recent results.\n\n"
f"## Your Task\n"
f"Provide helpful information about what the user might be looking for. "
f"If you have general knowledge about the topic, share it. "
f"Be honest that you don't have access to the very latest breaking news, but provide what context you can. "
f"Suggest that the user try checking major news websites directly or using a more specific search query.\n\n"
f"Provide a helpful response now:"
)
else:
prompt_for_llm = self._build_prompt_with_web(req, web_formatted)
llm_start = time.time()
llm_out = await self.llm.simple_call(prompt_for_llm, temperature=req.temperature)
llm_latency_ms = int((time.time() - llm_start) * 1000)
tools_used.append("llm")
estimated_tokens = len(llm_out) // 4 + len(prompt_for_llm) // 4
total_tokens += estimated_tokens
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="llm",
latency_ms=llm_latency_ms,
tokens_used=estimated_tokens,
success=True,
user_id=req.user_id
)
total_latency_ms = int((time.time() - start_time) * 1000)
self._analytics_log_agent_query(
tenant_id=req.tenant_id,
message_preview=req.message[:200],
intent=intent,
tools_used=tools_used,
total_tokens=total_tokens,
total_latency_ms=total_latency_ms,
success=True,
user_id=req.user_id
)
return AgentResponse(text=llm_out, decision=decision, tool_traces=tool_traces, reasoning_trace=reasoning_trace)
except Exception as web_err:
reasoning_trace.append({
"step": "web_search_forced_failed",
"error": str(web_err)[:200]
})
# Fall through to normal LLM path
# If the user is asking who the admin / owner is, try to ground the
# answer in tenant-specific RAG before falling back to a generic LLM reply.
user_text = req.message.lower()
# Normalize whitespace to make matching more robust
user_text_normalized = " ".join(user_text.split())
admin_phrases = [
"who is the admin",
"who's the admin",
"who is admin",
"who is the administrator",
"who's the administrator",
"who administers this platform",
"who administers the platform",
"who is the owner",
"who's the owner",
"who owns this platform",
"who owns the platform",
"who is the admin of integrachat",
"who's the admin of integrachat",
]
use_rag_for_admin = any(p in user_text_normalized for p in admin_phrases) or (
"admin" in user_text and "who" in user_text
)
prompt_for_llm = req.message
if use_rag_for_admin:
try:
rag_start = time.time()
rag_resp = await self.rag_with_repair(
query=req.message,
tenant_id=req.tenant_id,
original_threshold=0.2,
reasoning_trace=reasoning_trace,
user_id=req.user_id,
)
rag_latency_ms = int((time.time() - rag_start) * 1000)
tools_used.append("rag")
rag_formatted = self._format_tool_output("rag", rag_resp, rag_latency_ms)
tool_traces.append({"tool": "rag", "response": rag_formatted})
hits = self._extract_hits(rag_formatted)
hits_count = len(hits)
avg_score = rag_formatted.get("avg_score")
top_score = rag_formatted.get("top_score")
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="rag",
latency_ms=rag_latency_ms,
success=True,
user_id=req.user_id,
)
reasoning_trace.append(
{
"step": "tool_execution",
"tool": "rag",
"hit_count": hits_count,
"top_score": top_score,
"avg_score": avg_score,
"summary": self._summarize_hits(rag_formatted, limit=2),
"note": "admin_identity_override",
}
)
# For admin questions, answer directly from RAG and avoid any
# generic LLM behaviour. If there is at least one hit, return
# that snippet; otherwise return an explicit "don't know".
if hits:
best = hits[0]
admin_text = best.get("text") or best.get("content") or str(best)
llm_out = f"According to the tenant knowledge base, {admin_text.strip()}"
else:
llm_out = "I don't know who administers this platform based on the tenant data."
llm_latency_ms = 0
estimated_tokens = len(llm_out) // 4 + len(req.message) // 4
total_tokens += estimated_tokens
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="llm",
latency_ms=llm_latency_ms,
tokens_used=estimated_tokens,
success=True,
user_id=req.user_id,
)
reasoning_trace.append(
{
"step": "llm_response",
"mode": "admin_from_rag_only",
"latency_ms": llm_latency_ms,
"estimated_tokens": estimated_tokens,
}
)
total_latency_ms = int((time.time() - start_time) * 1000)
self._analytics_log_agent_query(
tenant_id=req.tenant_id,
message_preview=req.message[:200],
intent=intent,
tools_used=tools_used,
total_tokens=total_tokens,
total_latency_ms=total_latency_ms,
success=True,
user_id=req.user_id,
)
return AgentResponse(text=llm_out, decision=decision, reasoning_trace=reasoning_trace)
except Exception as rag_err:
reasoning_trace.append(
{
"step": "rag_for_admin_fallback",
"status": "error",
"error": str(rag_err),
}
)
# For all other questions, if we already have RAG hits from pgvector
# (rag_results from the prefetch step), reuse them to ground the
# LLM response instead of answering purely from the model.
# BUT: Skip RAG for news queries (they should use web search instead)
is_news_query_here = any(k in req.message.lower() for k in ["latest", "today", "news", "current", "recent", "breaking", "trending", "happening", "updates"])
news_patterns_here = [
r"latest news", r"current news", r"today's news", r"breaking news",
r"news about", r"news on", r"news of"
]
is_news_query_here = is_news_query_here or any(re.search(p, req.message.lower()) for p in news_patterns_here)
# NEVER use RAG for news queries - force web search or use general knowledge
if not use_rag_for_admin and rag_results and not is_news_query_here:
try:
rag_prefetched_dict: Dict[str, Any] = {"results": rag_results}
prompt_for_llm = self._build_prompt_with_rag(req, rag_prefetched_dict)
reasoning_trace.append(
{
"step": "rag_context_for_llm",
"hit_count": len(rag_results),
"note": "used_prefetched_pgvector_hits",
}
)
except Exception as build_err:
reasoning_trace.append(
{
"step": "rag_context_for_llm",
"status": "error",
"error": str(build_err),
}
)
elif not use_rag_for_admin:
# No RAG results available - enhance the prompt to still provide best answer
# BUT: For news queries, provide a helpful message about web search
if is_news_query_here:
prompt_for_llm = (
f"You are an assistant helping tenant {req.tenant_id}.\n\n"
f"## User Question\n{req.message}\n\n"
f"## Context\n"
f"The user is asking for latest news. I attempted to search for current information but didn't find specific results.\n\n"
f"## Your Task\n"
f"Provide helpful information about what the user might be looking for. "
f"If you have general knowledge about the topic, share it. "
f"Be honest that you don't have access to the very latest breaking news, but provide what context you can. "
f"Suggest that the user try checking major news websites directly or using a more specific search query.\n\n"
f"IMPORTANT: Do NOT say 'There is no mention of X in the provided context' - instead provide helpful general information or suggest where to find current news."
)
else:
prompt_for_llm = (
f"You are an assistant helping tenant {req.tenant_id}.\n\n"
f"## User Question\n{req.message}\n\n"
f"## Your Task\n"
f"Provide the best possible answer to the user's question. "
f"Be clear, accurate, comprehensive, and helpful. "
f"Focus on giving the user exactly what they need—clear guidance, accurate facts, "
f"and practical steps whenever possible. "
f"If you're uncertain about tenant-specific details, acknowledge that and provide general guidance."
)
llm_start = time.time()
llm_out = await self.llm.simple_call(prompt_for_llm, temperature=req.temperature)
llm_latency_ms = int((time.time() - llm_start) * 1000)
tools_used.append("llm")
estimated_tokens = len(llm_out) // 4 + len(prompt_for_llm) // 4
total_tokens += estimated_tokens
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="llm",
latency_ms=llm_latency_ms,
tokens_used=estimated_tokens,
success=True,
user_id=req.user_id
)
reasoning_trace.append({
"step": "llm_response",
"mode": "direct",
"latency_ms": llm_latency_ms,
"estimated_tokens": estimated_tokens
})
total_latency_ms = int((time.time() - start_time) * 1000)
self._analytics_log_agent_query(
tenant_id=req.tenant_id,
message_preview=req.message[:200],
intent=intent,
tools_used=tools_used,
total_tokens=total_tokens,
total_latency_ms=total_latency_ms,
success=True,
user_id=req.user_id
)
return AgentResponse(text=llm_out, decision=decision, reasoning_trace=reasoning_trace)
except Exception as e:
tool_traces.append({"tool": decision.tool, "error": str(e)})
try:
fallback = await self.llm.simple_call(req.message, temperature=req.temperature)
except Exception as llm_error:
error_msg = str(llm_error)
if "Groq API key" in error_msg or "GROQ_API_KEY" in error_msg:
fallback = (
f"I encountered an error while processing your request: {str(e)}\n\n"
f"Additionally, the AI service (Groq) is unavailable: {error_msg}\n\n"
f"To fix:\n"
f"1. Get a free Groq API key from https://console.groq.com\n"
f"2. Set GROQ_API_KEY in your .env file or environment variables"
)
else:
fallback = f"I encountered an error while processing your request: {str(e)}. Additionally, the AI service is unavailable: {error_msg}"
return AgentResponse(
text=fallback,
decision=AgentDecision(action="respond", tool=None, tool_input=None, reason=f"tool_error_fallback: {e}"),
tool_traces=tool_traces,
reasoning_trace=reasoning_trace + [{
"step": "error",
"tool": decision.tool,
"error": str(e)
}]
)
# Default: direct LLM response
# BUT: For news queries, try web search first even if tool selector didn't route to it
msg_lower = req.message.lower()
freshness_keywords = ["latest", "today", "news", "current", "recent",
"now", "updates", "breaking", "trending", "happening"]
news_patterns = [
r"latest news", r"current news", r"today's news", r"breaking news",
r"news about", r"news on", r"news of"
]
is_news_query_default = any(k in msg_lower for k in freshness_keywords) or \
any(re.search(p, msg_lower) for p in news_patterns)
# If it's a news query and we're in the default path, force web search
if is_news_query_default and decision.action != "call_tool" and decision.action != "multi_step":
try:
web_query = req.message
if len(web_query.split()) <= 4:
if "news" not in msg_lower:
web_query = f"{web_query} news latest"
web_start = time.time()
web_resp = await self.web_with_repair(
query=web_query,
tenant_id=req.tenant_id,
reasoning_trace=reasoning_trace,
user_id=req.user_id
)
web_latency_ms = int((time.time() - web_start) * 1000)
tools_used.append("web")
web_formatted = self._format_tool_output("web", web_resp, web_latency_ms)
# Context Engineering: Compress tool output if needed
web_formatted = await self.context_engineer.compressor.compress_tool_output("web", web_formatted)
tool_traces.append({"tool": "web", "response": web_formatted})
hits_count = len(self._extract_hits(web_formatted))
if hits_count > 0:
prompt = self._build_prompt_with_web(req, web_formatted)
else:
# Web search returned no results - use a news-specific prompt
prompt = (
f"You are an assistant helping tenant {req.tenant_id}.\n\n"
f"## User Question\n{req.message}\n\n"
f"## Context\n"
f"The user is asking for latest news, but web search did not return specific results for this query.\n\n"
f"## Your Task\n"
f"Provide helpful information about what the user might be looking for. "
f"If you know general information about the topic, share it. "
f"Be honest that you don't have access to the very latest news, but provide what context you can. "
f"Suggest that the user try rephrasing the query or checking news websites directly for the most current information."
)
llm_start = time.time()
llm_out = await self.llm.simple_call(prompt, temperature=req.temperature)
llm_latency_ms = int((time.time() - llm_start) * 1000)
tools_used.append("llm")
estimated_tokens = len(llm_out) // 4 + len(prompt) // 4
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="llm",
latency_ms=llm_latency_ms,
tokens_used=estimated_tokens,
success=True,
user_id=req.user_id
)
total_latency_ms = int((time.time() - start_time) * 1000)
self._analytics_log_agent_query(
tenant_id=req.tenant_id,
message_preview=req.message[:200],
intent=intent,
tools_used=tools_used,
total_tokens=estimated_tokens,
total_latency_ms=total_latency_ms,
success=True,
user_id=req.user_id
)
return AgentResponse(
text=llm_out,
decision=AgentDecision(action="respond", tool="web", tool_input=None, reason="news_query_forced_web_search"),
tool_traces=tool_traces,
reasoning_trace=reasoning_trace
)
except Exception as web_err:
# If web search fails, fall through to default LLM
reasoning_trace.append({
"step": "web_search_fallback",
"error": str(web_err)[:200]
})
try:
llm_start = time.time()
# For news queries in default path, use a better prompt
if is_news_query_default:
prompt_for_default = (
f"You are an assistant helping tenant {req.tenant_id}.\n\n"
f"## User Question\n{req.message}\n\n"
f"## Your Task\n"
f"The user is asking for latest news. I don't have access to real-time web search results right now. "
f"Please provide helpful information about what they might be looking for, or suggest they check news websites directly for the most current information."
)
else:
prompt_for_default = req.message
llm_out = await self.llm.simple_call(prompt_for_default, temperature=req.temperature)
llm_latency_ms = int((time.time() - llm_start) * 1000)
tools_used = ["llm"]
estimated_tokens = len(llm_out) // 4 + len(prompt_for_default) // 4
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="llm",
latency_ms=llm_latency_ms,
tokens_used=estimated_tokens,
success=True,
user_id=req.user_id
)
except Exception as e:
# If LLM fails, return a helpful error message
error_msg = str(e)
if "Groq API key" in error_msg or "GROQ_API_KEY" in error_msg:
llm_out = (
f"I couldn't connect to the AI service (Groq). "
f"Error: {error_msg}\n\n"
f"To fix this:\n"
f"1. Get a free Groq API key from https://console.groq.com\n"
f"2. Set GROQ_API_KEY in your .env file or environment variables\n"
f"3. Optionally set GROQ_MODEL (default: llama-3.1-8b-instant)"
)
else:
llm_out = f"I apologize, but I'm unable to process your request right now. The AI service is unavailable: {error_msg}"
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="llm",
success=False,
error_message=error_msg[:200],
user_id=req.user_id
)
reasoning_trace.append({
"step": "error",
"tool": "llm",
"error": str(e)
})
total_latency_ms = int((time.time() - start_time) * 1000)
self._analytics_log_agent_query(
tenant_id=req.tenant_id,
message_preview=req.message[:200],
intent=intent,
tools_used=tools_used if 'tools_used' in locals() else [],
total_tokens=estimated_tokens if 'estimated_tokens' in locals() else 0,
total_latency_ms=total_latency_ms,
success=True if 'llm_out' in locals() else False,
user_id=req.user_id
)
response = AgentResponse(
text=llm_out,
decision=AgentDecision(action="respond", tool=None, tool_input=None, reason="default_llm"),
reasoning_trace=reasoning_trace
)
# Cache successful response
self._cache_response(req, response)
return response
def _build_prompt_with_rag(self, req: AgentRequest, rag_resp: Dict[str, Any]) -> str:
snippets = []
scores_info = []
if isinstance(rag_resp, dict):
hits = rag_resp.get("results") or rag_resp.get("hits") or []
# Sort by score if available, take top 5
sorted_hits = sorted(
hits,
key=lambda h: float(h.get("score", h.get("similarity", 0.0))),
reverse=True
)[:5]
for i, h in enumerate(sorted_hits, 1):
txt = h.get("text") or h.get("content") or str(h)
score = h.get("score") or h.get("similarity", 0.0)
snippets.append(f"[Source {i}] {txt}")
scores_info.append(f"Source {i}: relevance score {score:.3f}")
snippet_text = "\n\n".join(snippets) if snippets else ""
scores_text = "\n".join(scores_info) if scores_info else ""
# Build optional relevance scores section separately to avoid complex f-string expressions
if scores_text:
relevance_section = "## Relevance Scores\n" + scores_text + "\n\n"
else:
relevance_section = ""
if not snippet_text:
prompt = (
f"You are an assistant helping tenant {req.tenant_id}.\n\n"
f"## User Question\n{req.message}\n\n"
f"## Context\n"
f"No relevant documents were found in the knowledge base for this question.\n\n"
f"## Important Rules\n"
f"If the user asks a question that cannot be answered directly from the Knowledge Base, "
f"then and ONLY then use the web-search tool to gather information. "
f"When using web search, keep the response short, factual, and neutral. "
f"Do NOT provide long legal, medical, or highly detailed professional explanations. "
f"If the topic involves legal, medical, financial, or safety-critical advice, provide a brief general explanation "
f"and tell the user to consult a qualified professional. "
f"Never present external information as part of the official Knowledge Base.\n\n"
f"## Your Task\n"
f"Since no Knowledge Base documents were found, you may use web search as a fallback if needed. "
f"Provide a brief, helpful answer. If you're uncertain about tenant-specific details, "
f"acknowledge that and provide general guidance. "
f"For legal, medical, financial, or safety-critical topics, keep responses brief and recommend consulting a professional."
)
else:
# Context Engineering: Get structured scratchpad context (Anthropic's note-taking)
scratchpad_context = self.context_engineer.get_scratchpad_context(limit=5)
scratchpad_section = f"\n## Structured Notes from Previous Steps\n{scratchpad_context}\n\n" if scratchpad_context else ""
# Build prompt with Anthropic's recommended structure
# Clear sections with XML/Markdown headers for better organization
prompt = (
f"\n"
f"You are an assistant helping tenant {req.tenant_id}. "
f"Your goal is to provide the most accurate, comprehensive, and helpful answer possible.\n"
f"\n\n"
f"\n"
f"## KB-First Strategy\n"
f"The Knowledge Base was checked first and relevant documents were found. "
f"Use these documents as your PRIMARY and AUTHORITATIVE source. "
f"Web search should ONLY be used as a fallback if the Knowledge Base cannot answer the question.\n"
f"{scratchpad_section}"
f"\n\n"
f"\n"
f"The following documents were retrieved from the tenant's knowledge base as relevant to the user's question:\n\n"
f"{snippet_text}\n\n"
f"{relevance_section}"
f"\n\n"
f"\n"
f"{req.message}\n"
f"\n\n"
f"\n"
f"## Your Task\n"
f"1. **Primary Goal**: Answer the user's question using the information from the knowledge base documents above.\n"
f"2. **KB Priority**: Base your answer PRIMARILY on the Knowledge Base. This is the authoritative source for tenant-specific information.\n"
f"3. **Accuracy**: Base your answer primarily on the highest-scoring sources (most relevant documents).\n"
f"4. **Comprehensiveness**: If multiple sources provide complementary information, synthesize them into a complete answer.\n"
f"5. **Citation**: When referencing specific information, indicate which source(s) you used (e.g., 'According to Source 1...' or 'Sources 1 and 2 indicate...').\n"
f"6. **Completeness**: If the documents don't fully answer the question, clearly state what information is available and what is missing.\n"
f"7. **Clarity**: Write in a clear, professional, and easy-to-understand manner.\n"
f"8. **Directness**: Get straight to the point - provide the answer the user needs without unnecessary preamble.\n"
f"\n\n"
f"Provide your answer now:"
)
return prompt
async def _execute_multi_step(self, req: AgentRequest, steps: List[Dict[str, Any]],
decision: AgentDecision, tool_traces: List[Dict[str, Any]],
reasoning_trace: List[Dict[str, Any]],
pre_fetched_rag: Optional[Dict[str, Any]] = None) -> AgentResponse:
"""
Execute multiple tools in sequence or parallel and synthesize results with LLM.
Supports parallel execution when steps are marked with "parallel" flag.
"""
start_time = time.time()
rag_data = None
web_data = None
admin_data = None
collected_data = []
tools_used = []
total_tokens = 0
# Detect if this is a news query - if so, skip RAG steps entirely
msg_lower = req.message.lower()
freshness_keywords = ["latest", "today", "news", "current", "recent",
"now", "updates", "breaking", "trending", "happening"]
news_patterns = [
r"latest news", r"current news", r"today's news", r"breaking news",
r"news about", r"news on", r"news of"
]
is_news_query = any(k in msg_lower for k in freshness_keywords) or \
any(re.search(p, msg_lower) for p in news_patterns)
# Filter out RAG steps for news queries
if is_news_query:
steps = [s for s in steps if s.get("tool") != "rag" and "rag" not in str(s.get("parallel", {}))]
reasoning_trace.append({
"step": "multi_step_news_filter",
"action": "removed_rag_steps",
"remaining_steps": [s.get("tool") if isinstance(s, dict) and "tool" in s else "parallel" for s in steps]
})
# Check if any step has parallel execution flag
parallel_step = None
for step_info in steps:
if step_info.get("parallel"):
parallel_step = step_info
break
# Handle parallel execution if detected
if parallel_step and parallel_step.get("parallel"):
parallel_config = parallel_step.get("parallel")
parallel_tasks = {}
start_time_parallel = time.time()
# Prepare parallel tasks with retry logic
# Skip RAG for news queries
if "rag" in parallel_config and not is_news_query:
rag_query = parallel_config["rag"]
if pre_fetched_rag:
# Use pre-fetched RAG if available - create a simple async function
async def get_prefetched_rag():
return pre_fetched_rag
parallel_tasks["rag"] = get_prefetched_rag()
else:
# Wrap with retry logic for parallel execution
async def rag_with_retry_wrapper():
return await self.rag_with_repair(
query=rag_query,
tenant_id=req.tenant_id,
original_threshold=0.3,
reasoning_trace=reasoning_trace,
user_id=req.user_id
)
parallel_tasks["rag"] = rag_with_retry_wrapper()
elif "rag" in parallel_config and is_news_query:
# Remove RAG from parallel config for news queries
parallel_config = {k: v for k, v in parallel_config.items() if k != "rag"}
reasoning_trace.append({
"step": "parallel_news_filter",
"action": "removed_rag_from_parallel",
"remaining_tools": list(parallel_config.keys())
})
if "web" in parallel_config:
web_query = parallel_config["web"]
# Wrap with retry logic for parallel execution
async def web_with_retry_wrapper():
return await self.web_with_repair(
query=web_query,
tenant_id=req.tenant_id,
reasoning_trace=reasoning_trace,
user_id=req.user_id
)
parallel_tasks["web"] = web_with_retry_wrapper()
# Execute tools in parallel
if parallel_tasks:
reasoning_trace.append({
"step": "parallel_execution",
"tools": list(parallel_tasks.keys()),
"mode": "parallel"
})
parallel_results = await self.run_parallel_tools(parallel_tasks)
parallel_latency_ms = int((time.time() - start_time_parallel) * 1000)
# Process RAG results
if "rag" in parallel_results:
rag_result = parallel_results["rag"]
if isinstance(rag_result, Exception):
tool_traces.append({"tool": "rag", "error": str(rag_result), "note": "parallel"})
reasoning_trace.append({
"step": "tool_execution",
"tool": "rag",
"status": "error",
"error": str(rag_result),
"latency_ms": parallel_latency_ms
})
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="rag",
latency_ms=parallel_latency_ms,
success=False,
error_message=str(rag_result)[:200],
user_id=req.user_id
)
else:
rag_data = rag_result
tools_used.append("rag")
tool_traces.append({"tool": "rag", "response": rag_result, "note": "parallel"})
hits_count = len(self._extract_hits(rag_result))
avg_score = None
top_score = None
if hits_count > 0:
scores = [h.get("score", 0.0) for h in self._extract_hits(rag_result) if isinstance(h, dict) and "score" in h]
if scores:
avg_score = sum(scores) / len(scores)
top_score = max(scores)
self._analytics_log_rag_search(
tenant_id=req.tenant_id,
query=req.message[:500],
hits_count=hits_count,
avg_score=avg_score,
top_score=top_score,
latency_ms=parallel_latency_ms
)
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="rag",
latency_ms=parallel_latency_ms,
success=True,
user_id=req.user_id
)
reasoning_trace.append({
"step": "tool_execution",
"tool": "rag",
"hit_count": hits_count,
"summary": self._summarize_hits(rag_result, limit=2),
"latency_ms": parallel_latency_ms,
"mode": "parallel"
})
# Process Web results
if "web" in parallel_results:
web_result = parallel_results["web"]
if isinstance(web_result, Exception):
tool_traces.append({"tool": "web", "error": str(web_result), "note": "parallel"})
reasoning_trace.append({
"step": "tool_execution",
"tool": "web",
"status": "error",
"error": str(web_result),
"latency_ms": parallel_latency_ms
})
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="web",
latency_ms=parallel_latency_ms,
success=False,
error_message=str(web_result)[:200],
user_id=req.user_id
)
else:
web_data = web_result
tools_used.append("web")
tool_traces.append({"tool": "web", "response": web_result, "note": "parallel"})
hits_count = len(self._extract_hits(web_result))
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="web",
latency_ms=parallel_latency_ms,
success=True,
user_id=req.user_id
)
reasoning_trace.append({
"step": "tool_execution",
"tool": "web",
"hit_count": hits_count,
"summary": self._summarize_hits(web_result, limit=2),
"latency_ms": parallel_latency_ms,
"mode": "parallel"
})
# Merge parallel results
merged_context = merge_parallel_results(parallel_results)
sources_list = list(set(e.get("source") for e in merged_context if e.get("source"))) if merged_context else []
reasoning_trace.append({
"step": "result_merger",
"merged_items": len(merged_context),
"sources": sources_list
})
# Format merged context for prompt
data_section = format_merged_context_for_prompt(merged_context, max_items=10)
else:
data_section = ""
else:
# Sequential execution (original logic)
parallel_tasks = {}
rag_parallel_query = self._first_query_for_tool(steps, "rag", req.message)
web_parallel_query = self._first_query_for_tool(steps, "web", req.message)
if rag_parallel_query and web_parallel_query and rag_parallel_query == web_parallel_query:
if not pre_fetched_rag:
parallel_tasks["rag"] = asyncio.create_task(self.mcp.call_rag(req.tenant_id, rag_parallel_query))
parallel_tasks["web"] = asyncio.create_task(self.mcp.call_web(req.tenant_id, web_parallel_query))
# Execute each step in sequence
for step_info in steps:
tool_name = step_info.get("tool")
step_input = step_info.get("input") or {}
query = step_input.get("query") or req.message
try:
if tool_name == "rag":
# Skip RAG for news queries
if is_news_query:
reasoning_trace.append({
"step": "tool_execution",
"tool": "rag",
"status": "skipped",
"reason": "news_query_detected"
})
continue # Skip this RAG step
# Reuse pre-fetched RAG if available, otherwise fetch with retry
if pre_fetched_rag and query == rag_parallel_query:
rag_resp = pre_fetched_rag
tool_traces.append({"tool": "rag", "response": rag_resp, "note": "used_pre_fetched"})
elif parallel_tasks.get("rag") and query == rag_parallel_query:
rag_resp = await parallel_tasks["rag"]
tool_traces.append({"tool": "rag", "response": rag_resp, "note": "parallel"})
else:
# Use autonomous retry with self-correction
rag_resp = await self.rag_with_repair(
query=query,
tenant_id=req.tenant_id,
original_threshold=0.3,
reasoning_trace=reasoning_trace,
user_id=req.user_id
)
tool_traces.append({"tool": "rag", "response": rag_resp, "note": "with_retry"})
rag_data = rag_resp
tools_used.append("rag")
hits = self._extract_hits(rag_resp)
reasoning_trace.append({
"step": "tool_execution",
"tool": "rag",
"hit_count": len(hits),
"summary": self._summarize_hits(rag_resp, limit=2)
})
# Extract snippets for prompt
if isinstance(rag_resp, dict):
for h in hits[:5]:
txt = h.get("text") or h.get("content") or str(h)
collected_data.append(f"[RAG] {txt}")
elif tool_name == "web":
if parallel_tasks.get("web") and query == web_parallel_query:
web_resp = await parallel_tasks["web"]
tool_traces.append({"tool": "web", "response": web_resp, "note": "parallel"})
else:
# Use autonomous retry with query rewriting
web_resp = await self.web_with_repair(
query=query,
tenant_id=req.tenant_id,
reasoning_trace=reasoning_trace,
user_id=req.user_id
)
tool_traces.append({"tool": "web", "response": web_resp, "note": "with_retry"})
web_data = web_resp
tools_used.append("web")
hits = self._extract_hits(web_resp)
reasoning_trace.append({
"step": "tool_execution",
"tool": "web",
"hit_count": len(hits),
"summary": self._summarize_hits(web_resp, limit=2)
})
# Extract snippets for prompt
if isinstance(web_resp, dict):
for h in hits[:5]:
title = h.get("title") or h.get("headline") or ""
snippet = h.get("snippet") or h.get("summary") or h.get("text") or ""
url = h.get("url") or h.get("link") or ""
collected_data.append(f"[WEB] {title}\n{snippet}\nSource: {url}")
elif tool_name == "admin":
admin_resp = await self.mcp.call_admin(req.tenant_id, query)
tool_traces.append({"tool": "admin", "response": admin_resp})
admin_data = admin_resp
tools_used.append("admin")
collected_data.append(f"[ADMIN] {json.dumps(admin_resp)}")
reasoning_trace.append({
"step": "tool_execution",
"tool": "admin",
"status": "completed"
})
elif tool_name == "llm":
# LLM is always last - synthesize all collected data
break
except Exception as e:
tool_traces.append({"tool": tool_name, "error": str(e)})
# Continue with other tools even if one fails
reasoning_trace.append({
"step": "error",
"tool": tool_name,
"error": str(e)
})
# Build comprehensive prompt with all collected data
data_section = "\n---\n".join(collected_data) if collected_data else ""
# Build final response. For admin-identity style questions, bypass generic
# multi-step LLM behaviour and answer directly from RAG data if available.
user_text = req.message.lower()
user_text_normalized = " ".join(user_text.split())
admin_phrases = [
"who is the admin",
"who's the admin",
"who is admin",
"who is the administrator",
"who's the administrator",
"who administers this platform",
"who administers the platform",
"who is the owner",
"who's the owner",
"who owns this platform",
"who owns the platform",
"who is the admin of integrachat",
"who's the admin of integrachat",
]
if any(p in user_text_normalized for p in admin_phrases) or ("admin" in user_text and "who" in user_text):
hits = self._extract_hits(rag_data) if rag_data else []
if hits:
best = hits[0]
admin_text = best.get("text") or best.get("content") or str(best)
llm_out = f"According to the tenant knowledge base, {admin_text.strip()}"
else:
llm_out = "I don't know who administers this platform based on the tenant data."
llm_latency_ms = 0
estimated_tokens = len(llm_out) // 4 + len(req.message) // 4
total_tokens += estimated_tokens
tools_used.append("llm")
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="llm",
latency_ms=llm_latency_ms,
tokens_used=estimated_tokens,
success=True,
user_id=req.user_id
)
total_latency_ms = int((time.time() - start_time) * 1000)
self._analytics_log_agent_query(
tenant_id=req.tenant_id,
message_preview=req.message[:200],
intent="multi_step",
tools_used=tools_used,
total_tokens=total_tokens,
total_latency_ms=total_latency_ms,
success=True,
user_id=req.user_id
)
return AgentResponse(
text=llm_out,
decision=decision,
tool_traces=tool_traces,
reasoning_trace=reasoning_trace + [{
"step": "llm_response",
"mode": "multi_step_admin_from_rag_only",
"latency_ms": llm_latency_ms,
"estimated_tokens": estimated_tokens
}]
)
# Otherwise, build the normal multi-step synthesis prompt.
if data_section:
# Check if we have both RAG and web data
has_rag = "[RAG]" in data_section
has_web = "[WEB]" in data_section
kb_first_note = ""
web_fallback_note = ""
if has_rag and has_web:
kb_first_note = (
f"\n## KB-First Strategy\n"
f"**Knowledge Base (RAG) was checked FIRST** and found relevant information. "
f"This is the PRIMARY and AUTHORITATIVE source. "
f"Web search results are provided as supplementary information only. "
f"Prioritize Knowledge Base information over web search results.\n\n"
)
web_fallback_note = (
f"\n## Web Search Rules\n"
f"When using web search information as supplementary data:\n"
f"- Keep web search details brief and factual\n"
f"- For legal, medical, financial, or safety topics, add: 'For specific advice, consult a qualified professional.'\n"
f"- Clearly distinguish between Knowledge Base (authoritative) and web search (supplementary) information\n\n"
)
elif has_web and not has_rag:
kb_first_note = (
f"\n## KB-First Strategy\n"
f"The Knowledge Base was checked FIRST but no relevant information was found. "
f"Web search results below are provided as a FALLBACK. "
f"Keep the response short, factual, and neutral. "
f"For legal, medical, financial, or safety topics, recommend consulting a qualified professional.\n\n"
)
# Get structured scratchpad context (Anthropic's note-taking)
scratchpad_context = self.context_engineer.get_scratchpad_context(limit=5)
scratchpad_section = f"\n## Structured Notes\n{scratchpad_context}\n" if scratchpad_context else ""
# Build prompt with Anthropic's structured format (XML-style sections)
# Pre-build optional web guidance section to avoid complex nested f-strings
if web_fallback_note:
web_guidance_section = (
"\n"
f"{web_fallback_note.strip()}\n"
"\n\n"
)
else:
web_guidance_section = ""
if has_web:
brief_web_instruction = (
"11. **Brief Web Content**: If using web search, keep that portion of the response brief "
"(2-4 sentences). Add professional disclaimers for legal/medical/financial topics.\n"
)
else:
brief_web_instruction = ""
prompt = (
f"\n"
f"You are an assistant helping tenant {req.tenant_id}. "
f"Your goal is to provide the most accurate, comprehensive, and helpful answer possible.\n"
f"\n\n"
f"\n"
f"{kb_first_note.strip()}"
f"{scratchpad_section}"
f"\n\n"
f"\n"
f"The following details have been gathered from reliable sources:\n\n"
f"{data_section}\n"
f"\n\n"
f"{web_guidance_section}"
f"\n"
f"{req.message}\n"
f"\n\n"
f"\n"
f"## Your Task\n"
f"1. **Primary Goal**: Use the information above to directly and completely address the user's request.\n"
f"2. **Source Priority**: {'If both Knowledge Base (RAG) and web search results are present, prioritize Knowledge Base as the authoritative source. ' if has_rag and has_web else ''}Use web search information only to supplement or when KB has no relevant information.\n"
f"3. **Synthesis**: Combine information from different sources when they provide complementary details.\n"
f"4. **Prioritization**: If sources conflict, prioritize Knowledge Base information over web search results.\n"
f"5. **Completeness**: Provide a comprehensive answer that covers all aspects of the user's question.\n"
f"6. **Accuracy**: Base your answer on the provided information. If information is missing or uncertain, clearly state that.\n"
f"7. **Clarity**: Write in a clear, professional, and easy-to-understand manner.\n"
f"8. **Directness**: Get straight to the point - provide the answer the user needs without unnecessary preamble.\n"
f"9. **Actionability**: If the question requires steps or actions, provide clear, actionable guidance.\n"
f"10. **Citation**: When referencing specific sources, indicate which source(s) you used (e.g., '[RAG]', '[WEB]').\n"
f"{brief_web_instruction}"
f"\n\n"
f"If the information is incomplete, explain what can and cannot be concluded from the available data. "
f"Focus on giving the user exactly what they need—clear guidance, accurate facts, and practical steps whenever possible.\n\n"
f"Provide your comprehensive answer now:"
)
else:
# No data collected, provide best answer from general knowledge
prompt = (
f"You are an assistant helping tenant {req.tenant_id}.\n\n"
f"## User Question\n{req.message}\n\n"
f"## Context\n"
f"No specific information was found in the knowledge base or web search for this question.\n\n"
f"## Your Task\n"
f"Provide the best possible answer based on your general knowledge. "
f"Be clear, accurate, comprehensive, and helpful. "
f"If you're uncertain about tenant-specific details, acknowledge that and provide general guidance. "
f"Focus on giving the user exactly what they need—clear guidance, accurate facts, and practical steps whenever possible."
)
# Final LLM synthesis
try:
llm_start = time.time()
llm_out = await self.llm.simple_call(prompt, temperature=req.temperature)
llm_latency_ms = int((time.time() - llm_start) * 1000)
tools_used.append("llm")
estimated_tokens = len(llm_out) // 4 + len(prompt) // 4
total_tokens += estimated_tokens
self._analytics_log_tool_usage(
tenant_id=req.tenant_id,
tool_name="llm",
latency_ms=llm_latency_ms,
tokens_used=estimated_tokens,
success=True,
user_id=req.user_id
)
total_latency_ms = int((time.time() - start_time) * 1000)
self._analytics_log_agent_query(
tenant_id=req.tenant_id,
message_preview=req.message[:200],
intent="multi_step",
tools_used=tools_used,
total_tokens=total_tokens,
total_latency_ms=total_latency_ms,
success=True,
user_id=req.user_id
)
return AgentResponse(
text=llm_out,
decision=decision,
tool_traces=tool_traces,
reasoning_trace=reasoning_trace + [{
"step": "llm_response",
"mode": "multi_step_parallel" if parallel_step else "multi_step",
"latency_ms": llm_latency_ms,
"estimated_tokens": estimated_tokens
}]
)
except Exception as e:
tool_traces.append({"tool": "llm", "error": str(e)})
error_msg = str(e)
# Provide helpful error message
if "Groq API key" in error_msg or "GROQ_API_KEY" in error_msg:
fallback = (
f"I couldn't connect to the AI service (Groq). "
f"Error: {error_msg}\n\n"
f"To fix this:\n"
f"1. Get a free Groq API key from https://console.groq.com\n"
f"2. Set GROQ_API_KEY in your .env file or environment variables\n"
f"3. Optionally set GROQ_MODEL (default: llama-3.1-8b-instant)"
)
else:
fallback = f"I encountered an error while synthesizing the response: {error_msg}"
return AgentResponse(
text=fallback,
decision=AgentDecision(
action="respond",
tool=None,
tool_input=None,
reason=f"multi_step_llm_error: {e}"
),
tool_traces=tool_traces,
reasoning_trace=reasoning_trace + [{
"step": "error",
"tool": "llm",
"error": str(e)
}]
)
# =============================================================
# AUTONOMOUS RETRY + SELF-CORRECTION SYSTEM
# =============================================================
"""
This system provides autonomous retry and self-correction capabilities
for the agent orchestrator. It enables the agent to:
1. **Self-healing**: Tools that break automatically retry with adjusted parameters
2. **Resilient operations**: Handles low RAG scores, empty web results, and misfired rules
3. **Smart optimization**: Automatically rewrites queries, adjusts thresholds, and optimizes parameters
4. **Enterprise-grade reliability**: Matches enterprise behavior with comprehensive retry strategies
Key features:
- safe_tool_call(): Generic retry wrapper for any tool call
- rag_with_repair(): RAG search with automatic threshold adjustment and query expansion
- web_with_repair(): Web search with automatic query rewriting for empty results
- rule_safe_message(): Message rewriting to comply with admin rules
All retry attempts are logged to analytics for monitoring and debugging.
"""
async def safe_tool_call(
self,
tool_fn,
params: Dict[str, Any],
max_retries: int = 2,
fallback_params: Optional[Dict[str, Any]] = None,
tool_name: str = "unknown",
tenant_id: Optional[str] = None,
user_id: Optional[str] = None,
reasoning_trace: Optional[List[Dict[str, Any]]] = None
) -> Dict[str, Any]:
"""
Wrapper for tool calls with automatic retry and self-correction.
Args:
tool_fn: Async function to call
params: Parameters to pass to tool_fn
max_retries: Maximum number of retry attempts
fallback_params: Alternative parameters to try if initial attempt fails
tool_name: Name of the tool (for logging)
tenant_id: Tenant ID (for analytics)
user_id: User ID (for analytics)
reasoning_trace: Optional reasoning trace to append to
Returns:
Tool result dictionary, or {"error": "tool_failed_after_retries"} if all attempts fail
"""
for attempt in range(max_retries):
try:
result = await tool_fn(**params)
if attempt > 0:
# Log successful retry
if reasoning_trace is not None:
reasoning_trace.append({
"step": "retry_success",
"tool": tool_name,
"attempt": attempt + 1,
"status": "recovered"
})
if tenant_id:
self._analytics_log_tool_usage(
tenant_id=tenant_id,
tool_name=f"{tool_name}_retry_{attempt+1}",
latency_ms=0,
success=True,
user_id=user_id
)
return result
except Exception as e:
error_msg = str(e)
if reasoning_trace is not None:
reasoning_trace.append({
"step": "retry_attempt",
"tool": tool_name,
"attempt": attempt + 1,
"error": error_msg[:200]
})
# Log failed attempt
if tenant_id:
self._analytics_log_tool_usage(
tenant_id=tenant_id,
tool_name=tool_name,
latency_ms=0,
success=False,
error_message=error_msg[:200],
user_id=user_id
)
# Try alternate params if provided and not last attempt
if fallback_params and attempt < max_retries - 1:
params = {**params, **fallback_params}
if reasoning_trace is not None:
reasoning_trace.append({
"step": "retry_with_fallback_params",
"tool": tool_name,
"attempt": attempt + 2,
"fallback_params": fallback_params
})
# If last attempt, return error
if attempt == max_retries - 1:
return {"error": "tool_failed_after_retries", "error_message": error_msg}
return {"error": "tool_failed_after_retries"}
async def rag_with_repair(
self,
query: str,
tenant_id: str,
original_threshold: float = 0.3,
reasoning_trace: Optional[List[Dict[str, Any]]] = None,
user_id: Optional[str] = None
) -> Dict[str, Any]:
"""
RAG search with automatic self-correction for low scores.
Strategy:
1. Try with original threshold
2. If top_score < 0.30, retry with lower threshold (0.15)
3. If still low (< 0.15), expand query and retry
"""
# Initial attempt
rag_start = time.time()
result = await self.mcp.call_rag(tenant_id, query, threshold=original_threshold)
rag_latency_ms = int((time.time() - rag_start) * 1000)
# Extract hits and calculate scores
hits = self._extract_hits(result)
top_score = None
avg_score = None
if hits:
scores = [h.get("score", 0.0) for h in hits if isinstance(h, dict) and "score" in h]
if scores:
top_score = max(scores)
avg_score = sum(scores) / len(scores)
if reasoning_trace is not None:
reasoning_trace.append({
"step": "rag_initial_search",
"query": query[:200],
"hits_count": len(hits),
"top_score": top_score,
"avg_score": avg_score,
"threshold": original_threshold
})
# Retry logic: low score → lower threshold
if top_score is not None and top_score < 0.30 and original_threshold >= 0.15:
if reasoning_trace is not None:
reasoning_trace.append({
"step": "rag_retry_low_threshold",
"reason": f"top_score {top_score:.3f} < 0.30, retrying with threshold=0.15"
})
retry_start = time.time()
result = await self.mcp.call_rag(tenant_id, query, threshold=0.15)
retry_latency_ms = int((time.time() - retry_start) * 1000)
rag_latency_ms += retry_latency_ms
hits = self._extract_hits(result)
if hits:
scores = [h.get("score", 0.0) for h in hits if isinstance(h, dict) and "score" in h]
if scores:
top_score = max(scores)
avg_score = sum(scores) / len(scores)
# Log retry
self._analytics_log_tool_usage(
tenant_id=tenant_id,
tool_name="rag_retry_low_threshold",
latency_ms=retry_latency_ms,
success=True,
user_id=user_id
)
# Final retry: expand query if score still too low
if top_score is not None and top_score < 0.15:
expanded_query = f"{query} (more details comprehensive explanation)"
if reasoning_trace is not None:
reasoning_trace.append({
"step": "rag_retry_expanded_query",
"reason": f"top_score {top_score:.3f} < 0.15, retrying with expanded query",
"original_query": query[:200],
"expanded_query": expanded_query[:200]
})
retry_start = time.time()
result = await self.mcp.call_rag(tenant_id, expanded_query, threshold=0.15)
retry_latency_ms = int((time.time() - retry_start) * 1000)
rag_latency_ms += retry_latency_ms
hits = self._extract_hits(result)
if hits:
scores = [h.get("score", 0.0) for h in hits if isinstance(h, dict) and "score" in h]
if scores:
top_score = max(scores)
avg_score = sum(scores) / len(scores)
# Log retry
self._analytics_log_tool_usage(
tenant_id=tenant_id,
tool_name="rag_retry_expanded_query",
latency_ms=retry_latency_ms,
success=True,
user_id=user_id
)
if reasoning_trace is not None:
reasoning_trace.append({
"step": "rag_expanded_query_result",
"hits_count": len(hits),
"top_score": top_score,
"avg_score": avg_score
})
# Log final RAG search
if hits:
self._analytics_log_rag_search(
tenant_id=tenant_id,
query=query[:500],
hits_count=len(hits),
avg_score=avg_score,
top_score=top_score,
latency_ms=rag_latency_ms
)
return result
async def web_with_repair(
self,
query: str,
tenant_id: str,
reasoning_trace: Optional[List[Dict[str, Any]]] = None,
user_id: Optional[str] = None
) -> Dict[str, Any]:
"""
Web search with multi-query strategy and automatic query rewriting.
Strategy:
1. Try original query
2. If empty, generate multiple query variations using query expander
3. Execute queries in parallel for better results
4. Merge results from all successful queries
"""
# Detect if this is a news query
query_lower = query.lower()
is_news_query = any(kw in query_lower for kw in ["news", "latest", "breaking", "current", "today", "recent", "update"])
# Initial attempt
web_start = time.time()
result = await self.mcp.call_web(tenant_id, query)
web_latency_ms = int((time.time() - web_start) * 1000)
hits = self._extract_hits(result)
if reasoning_trace is not None:
reasoning_trace.append({
"step": "web_initial_search",
"query": query[:200],
"hits_count": len(hits),
"is_news_query": is_news_query
})
# Multi-query strategy: if initial results are poor, try multiple variations in parallel
if not result or len(hits) < 3:
# Generate query variations
if is_news_query:
# Use query expander for news queries
try:
query_variations = self.query_expander.expand_news_query(query)
except Exception:
query_variations = [
f"{query} news",
f"latest {query}",
f"{query} latest news",
f"breaking news {query}"
]
else:
# For general queries, try explanation-focused rewrites
query_variations = [
f"best explanation of {query}",
f"{query} facts summary",
f"information about {query}",
f"what is {query}"
]
# Execute multiple queries in parallel
if len(query_variations) > 1:
async def search_variation(q: str):
try:
return await self.mcp.call_web(tenant_id, q)
except Exception as e:
logger.debug(f"Web search failed for query '{q}': {e}")
return None
# Run all variations in parallel
parallel_tasks = {q: search_variation(q) for q in query_variations[:3]} # Limit to 3 parallel
parallel_results = await self.run_parallel_tools(parallel_tasks)
# Merge results from all successful queries
all_hits = []
seen_urls = set()
# Add original hits
for hit in hits:
url = hit.get("url") or hit.get("link", "")
if url and url not in seen_urls:
all_hits.append(hit)
seen_urls.add(url)
# Add hits from parallel queries
for q, res in parallel_results.items():
if res and not isinstance(res, Exception):
var_hits = self._extract_hits(res)
for hit in var_hits:
url = hit.get("url") or hit.get("link", "")
if url and url not in seen_urls:
all_hits.append(hit)
seen_urls.add(url)
# Update result with merged hits
if all_hits:
result = {"results": all_hits[:10]} # Limit to top 10
hits = all_hits[:10]
if reasoning_trace is not None:
reasoning_trace.append({
"step": "web_multi_query_merge",
"variations_tried": len(query_variations),
"total_hits_merged": len(all_hits),
"final_hits_count": len(hits)
})
# If parallel didn't help, try one more sequential attempt with best variation
if not all_hits and len(query_variations) > 0:
best_variation = query_variations[0]
retry_start = time.time()
try:
result = await self.mcp.call_web(tenant_id, best_variation)
retry_latency_ms = int((time.time() - retry_start) * 1000)
web_latency_ms += retry_latency_ms
hits = self._extract_hits(result)
if hits:
if reasoning_trace is not None:
reasoning_trace.append({
"step": "web_sequential_fallback_success",
"query": best_variation[:200],
"hits_count": len(hits)
})
except Exception as e:
logger.debug(f"Final web search retry failed: {e}")
# Log final web search
self._analytics_log_tool_usage(
tenant_id=tenant_id,
tool_name="web",
latency_ms=web_latency_ms,
success=len(hits) > 0,
user_id=user_id
)
return result
async def rule_safe_message(
self,
user_message: str,
tenant_id: str,
reasoning_trace: Optional[List[Dict[str, Any]]] = None
) -> str:
"""
Check admin rules and rewrite message if it violates policies.
Strategy:
1. Check rules
2. If blocked, ask LLM to rewrite message to comply
3. Return safe version
"""
matches: List[RedFlagMatch] = await self.redflag.check(tenant_id, user_message)
if not matches:
return user_message
# Check if any are blocking rules (not just brief response rules)
blocking_rules = []
for match in matches:
rule_text = (match.description or match.pattern or "").lower()
is_brief_rule = (
match.severity == "low" and (
"greeting" in rule_text or
"brief" in rule_text or
"simple response" in rule_text
)
)
if not is_brief_rule:
blocking_rules.append(match)
# Only rewrite if there are blocking rules
if not blocking_rules:
return user_message
if reasoning_trace is not None:
reasoning_trace.append({
"step": "rule_violation_detected",
"blocking_rules_count": len(blocking_rules),
"action": "attempting_rewrite"
})
# Ask LLM to rewrite the message
rewrite_prompt = f"""The following user message violates company policies. Rewrite it to be compliant while preserving the user's intent as much as possible.
Original message: "{user_message}"
Violated policies:
{chr(10).join([f"- {m.description or m.pattern}" for m in blocking_rules[:3]])}
Provide a rewritten version that:
1. Avoids the policy violations
2. Preserves the user's original intent
3. Remains professional and helpful
Rewritten message:"""
try:
rewritten = await self.llm.simple_call(rewrite_prompt, temperature=0.3)
rewritten = rewritten.strip().strip('"').strip("'")
if reasoning_trace is not None:
reasoning_trace.append({
"step": "rule_rewrite_completed",
"original_length": len(user_message),
"rewritten_length": len(rewritten),
"rewritten_preview": rewritten[:200]
})
# Verify the rewritten message doesn't trigger rules
verify_matches = await self.redflag.check(tenant_id, rewritten)
if not verify_matches or all(
(m.description or m.pattern or "").lower() in ["greeting", "brief", "simple response"]
for m in verify_matches
):
return rewritten
if reasoning_trace is not None:
reasoning_trace.append({
"step": "rule_rewrite_still_violates",
"action": "using_original_with_block"
})
except Exception as e:
if reasoning_trace is not None:
reasoning_trace.append({
"step": "rule_rewrite_failed",
"error": str(e)[:200]
})
# Return original if rewrite failed or still violates
return user_message
def _build_prompt_with_web(self, req: AgentRequest, web_resp: Dict[str, Any]) -> str:
snippets = []
if isinstance(web_resp, dict):
hits = web_resp.get("results") or web_resp.get("items") or []
for i, h in enumerate(hits[:5], 1):
title = h.get("title") or h.get("headline") or ""
snippet = h.get("snippet") or h.get("summary") or h.get("text") or ""
url = h.get("url") or h.get("link") or ""
display_link = h.get("displayLink") or h.get("display_link") or ""
source_info = display_link if display_link else url.split('/')[2] if url else "Unknown source"
snippets.append(f"[Result {i}] {title}\n{snippet}\nSource: {source_info} ({url})")
snippet_text = "\n\n".join(snippets) if snippets else ""
if not snippet_text:
prompt = (
f"You are an assistant helping tenant {req.tenant_id}.\n\n"
f"## User Question\n{req.message}\n\n"
f"## Context\n"
f"No relevant web search results were found for this question.\n\n"
f"## Your Task\n"
f"Provide the best possible answer based on your general knowledge. "
f"Be clear, accurate, and helpful. If you're uncertain about specific details, "
f"acknowledge that and provide general guidance."
)
else:
# Build prompt with Anthropic's recommended structure (clear sections with XML tags)
prompt = (
f"\n"
f"You are an assistant helping tenant {req.tenant_id}. "
f"The Knowledge Base was checked first but no relevant information was found. "
f"Web search results are provided below as a fallback.\n"
f"\n\n"
f"\n"
f"## Important Rules for Web Search Responses\n"
f"1. **KB-First Approach**: Always check Knowledge Base first. Web search is ONLY a fallback when KB has no relevant information.\n"
f"2. **Keep it Short**: When using web search, keep responses short, factual, and neutral. Do NOT provide long explanations.\n"
f"3. **No Professional Advice**: Do NOT provide long legal, medical, or highly detailed professional explanations. "
f"If the topic involves legal, medical, financial, or safety-critical advice, provide a brief general explanation "
f"and tell the user to consult a qualified professional.\n"
f"4. **Clear Source Distinction**: Never present external web search information as part of the official Knowledge Base. "
f"Always clarify that this information comes from external sources.\n"
f"5. **Safety First**: For safety-critical topics, always recommend consulting qualified professionals.\n"
f"\n\n"
f"\n"
f"The following search results were found for the user's question:\n\n"
f"{snippet_text}\n"
f"\n\n"
f"\n"
f"{req.message}\n"
f"\n\n"
f"\n"
f"## Your Task\n"
f"1. **Primary Goal**: Provide a short, factual answer using the web search results above.\n"
f"2. **Keep it Brief**: Limit your response to 2-4 sentences. Do NOT provide lengthy explanations.\n"
f"3. **Accuracy**: Prioritize information from authoritative sources (recognized websites, official sources, etc.).\n"
f"4. **Professional Disclaimers**: For legal, medical, financial, or safety topics, include: "
f"'For specific advice, please consult a qualified professional.'\n"
f"5. **Source Clarity**: Start by mentioning this information comes from web search, not the Knowledge Base.\n"
f"6. **Citation**: Briefly indicate which source(s) you used.\n"
f"\n\n"
f"Provide a short, helpful answer now:"
)
return prompt
def _format_tool_output(self, tool_name: str, output: Any, latency_ms: int) -> Dict[str, Any]:
"""
Format tool output to conform to strict JSON schema.
Args:
tool_name: Name of the tool (rag, web, admin, llm)
output: Raw tool output
latency_ms: Actual latency in milliseconds
Returns:
Formatted output conforming to tool schema
"""
if tool_name == "rag":
# Format RAG output
if isinstance(output, dict):
results = output.get("results") or output.get("hits") or []
# Ensure each result has required fields
formatted_results = []
for r in results:
if isinstance(r, dict):
formatted_results.append({
"text": r.get("text") or r.get("content") or str(r),
"similarity": float(r.get("similarity") or r.get("score") or 0.0),
"metadata": r.get("metadata") or {},
"doc_id": r.get("doc_id") or r.get("id")
})
else:
formatted_results.append({
"text": str(r),
"similarity": 0.5,
"metadata": {},
"doc_id": None
})
# Calculate aggregate scores
scores = [r["similarity"] for r in formatted_results if r["similarity"] > 0]
avg_score = sum(scores) / len(scores) if scores else 0.0
top_score = max(scores) if scores else 0.0
return {
"results": formatted_results,
"query": output.get("query", ""),
"tenant_id": output.get("tenant_id", ""),
"hits_count": len(formatted_results),
"avg_score": round(avg_score, 3),
"top_score": round(top_score, 3),
"latency_ms": latency_ms
}
else:
# Fallback for non-dict output
return {
"results": [{"text": str(output), "similarity": 0.5, "metadata": {}, "doc_id": None}],
"query": "",
"tenant_id": "",
"hits_count": 1,
"avg_score": 0.5,
"top_score": 0.5,
"latency_ms": latency_ms
}
elif tool_name == "web":
# Format Web output
if isinstance(output, dict):
results = output.get("results") or output.get("items") or []
formatted_results = []
for r in results:
if isinstance(r, dict):
formatted_results.append({
"title": r.get("title") or r.get("headline") or "",
"snippet": r.get("snippet") or r.get("summary") or r.get("text") or "",
"link": r.get("url") or r.get("link") or "",
"displayLink": r.get("displayLink") or r.get("display_link") or ""
})
else:
formatted_results.append({
"title": "",
"snippet": str(r),
"link": "",
"displayLink": ""
})
return {
"results": formatted_results,
"query": output.get("query", ""),
"total_results": output.get("total_results") or output.get("totalResults") or len(formatted_results),
"latency_ms": latency_ms
}
else:
return {
"results": [],
"query": "",
"total_results": 0,
"latency_ms": latency_ms
}
elif tool_name == "admin":
# Format Admin output
if isinstance(output, dict):
violations = output.get("violations") or output.get("matches") or []
formatted_violations = []
for v in violations:
if isinstance(v, dict):
formatted_violations.append({
"rule_id": v.get("rule_id") or v.get("id") or "",
"rule_pattern": v.get("rule_pattern") or v.get("pattern") or "",
"severity": v.get("severity", "medium"),
"matched_text": v.get("matched_text") or v.get("text") or "",
"confidence": float(v.get("confidence", 1.0)),
"message_preview": v.get("message_preview") or v.get("preview") or ""
})
return {
"violations": formatted_violations,
"checked": output.get("checked", True),
"rules_count": output.get("rules_count") or output.get("rulesCount") or len(formatted_violations),
"latency_ms": latency_ms
}
else:
return {
"violations": [],
"checked": True,
"rules_count": 0,
"latency_ms": latency_ms
}
elif tool_name == "llm":
# Format LLM output
if isinstance(output, str):
return {
"text": output,
"tokens_used": len(output) // 4, # Rough estimate
"latency_ms": latency_ms,
"model": getattr(self.llm, 'model', 'unknown'),
"temperature": 0.0
}
elif isinstance(output, dict):
return {
"text": output.get("text") or output.get("response") or str(output),
"tokens_used": output.get("tokens_used") or output.get("tokens") or 0,
"latency_ms": latency_ms,
"model": output.get("model") or getattr(self.llm, 'model', 'unknown'),
"temperature": output.get("temperature", 0.0)
}
else:
return {
"text": str(output),
"tokens_used": 0,
"latency_ms": latency_ms,
"model": getattr(self.llm, 'model', 'unknown'),
"temperature": 0.0
}
# Unknown tool - return as-is
return output if isinstance(output, dict) else {"output": str(output), "latency_ms": latency_ms}
@staticmethod
def _extract_hits(resp: Optional[Dict[str, Any]]) -> List[Dict[str, Any]]:
if not isinstance(resp, dict):
return []
return resp.get("results") or resp.get("hits") or resp.get("items") or []
def _summarize_hits(self, resp: Optional[Dict[str, Any]], limit: int = 3) -> List[str]:
hits = self._extract_hits(resp)
summaries = []
for hit in hits[:limit]:
if isinstance(hit, dict):
snippet = hit.get("text") or hit.get("content") or hit.get("snippet") or ""
else:
snippet = str(hit)
summaries.append(snippet[:160])
return summaries
async def run_parallel_tools(self, tasks: Dict[str, Any]) -> Dict[str, Any]:
"""
Run multiple tools in parallel using asyncio.gather.
Args:
tasks: Dictionary mapping tool names to coroutines, e.g.:
{"rag": rag_coro, "web": web_coro}
Returns:
Dictionary mapping tool names to results, e.g.:
{"rag": rag_result, "web": web_result}
Exceptions are returned as values if a tool fails.
"""
if not tasks:
return {}
names = list(tasks.keys())
coros = [tasks[name] for name in names]
# Run all coroutines in parallel, return exceptions instead of raising
results = await asyncio.gather(*coros, return_exceptions=True)
return {names[i]: results[i] for i in range(len(names))}
@staticmethod
def _first_query_for_tool(steps: List[Dict[str, Any]], tool_name: str, default_query: str) -> Optional[str]:
for step in steps:
if step.get("tool") == tool_name:
input_data = step.get("input") or {}
return input_data.get("query") or default_query
return None