SPARKNET / src /rag /agentic /orchestrator.py
MHamdan's picture
Initial commit: SPARKNET framework
d520909
"""
Agentic RAG Orchestrator
Coordinates the multi-agent RAG pipeline with self-correction loop.
Follows FAANG best practices for production RAG systems.
Pipeline:
Query -> Plan -> Retrieve -> Rerank -> Synthesize -> Validate -> (Revise?) -> Response
Key Features:
- LangGraph-style state machine
- Self-correction loop (up to N attempts)
- Streaming support
- Comprehensive logging and metrics
- Graceful degradation
"""
from typing import List, Optional, Dict, Any, Generator, Tuple
from pydantic import BaseModel, Field
from loguru import logger
from dataclasses import dataclass, field
from enum import Enum
import time
from ..store import VectorStore, get_vector_store, VectorStoreConfig
from ..embeddings import EmbeddingAdapter, get_embedding_adapter, EmbeddingConfig
from .query_planner import QueryPlannerAgent, QueryPlan, SubQuery
from .retriever import RetrieverAgent, RetrievalResult, HybridSearchConfig
from .reranker import RerankerAgent, RankedResult, RerankerConfig
from .synthesizer import SynthesizerAgent, SynthesisResult, Citation, SynthesizerConfig
from .critic import CriticAgent, CriticResult, ValidationIssue, CriticConfig
class PipelineStage(str, Enum):
"""Stages in the RAG pipeline."""
PLANNING = "planning"
RETRIEVAL = "retrieval"
RERANKING = "reranking"
SYNTHESIS = "synthesis"
VALIDATION = "validation"
REVISION = "revision"
COMPLETE = "complete"
class RAGConfig(BaseModel):
"""Configuration for the agentic RAG system."""
# LLM settings (shared across agents)
model: str = Field(default="llama3.2:3b")
base_url: str = Field(default="http://localhost:11434")
# Pipeline settings
max_revision_attempts: int = Field(default=2, ge=0, le=5)
enable_query_planning: bool = Field(default=True)
enable_reranking: bool = Field(default=True)
enable_validation: bool = Field(default=True)
# Retrieval settings
retrieval_top_k: int = Field(default=10, ge=1)
final_top_k: int = Field(default=5, ge=1)
# Confidence thresholds
min_confidence: float = Field(default=0.5, ge=0.0, le=1.0)
# Logging
verbose: bool = Field(default=False)
@dataclass
class RAGState:
"""State maintained through the pipeline."""
query: str
stage: PipelineStage = PipelineStage.PLANNING
# Intermediate results
query_plan: Optional[QueryPlan] = None
retrieved_chunks: List[RetrievalResult] = field(default_factory=list)
ranked_chunks: List[RankedResult] = field(default_factory=list)
synthesis_result: Optional[SynthesisResult] = None
critic_result: Optional[CriticResult] = None
# Revision tracking
revision_attempt: int = 0
revision_history: List[SynthesisResult] = field(default_factory=list)
# Metrics
start_time: float = field(default_factory=time.time)
stage_times: Dict[str, float] = field(default_factory=dict)
# Errors
errors: List[str] = field(default_factory=list)
class RAGResponse(BaseModel):
"""Final response from the RAG system."""
answer: str
citations: List[Citation]
confidence: float
# Metadata
query: str
num_sources: int
validated: bool
revision_attempts: int
# Detailed info (optional)
query_plan: Optional[Dict[str, Any]] = None
validation_details: Optional[Dict[str, Any]] = None
latency_ms: float = 0.0
class AgenticRAG:
"""
Production-grade Multi-Agent RAG System.
Orchestrates:
- QueryPlannerAgent: Query decomposition and planning
- RetrieverAgent: Hybrid retrieval
- RerankerAgent: Cross-encoder reranking
- SynthesizerAgent: Answer generation
- CriticAgent: Validation and hallucination detection
Features:
- Self-correction loop
- Graceful degradation
- Comprehensive metrics
"""
def __init__(
self,
config: Optional[RAGConfig] = None,
vector_store: Optional[VectorStore] = None,
embedding_adapter: Optional[EmbeddingAdapter] = None,
):
"""
Initialize the Agentic RAG system.
Args:
config: RAG configuration
vector_store: Vector store for retrieval
embedding_adapter: Embedding adapter
"""
self.config = config or RAGConfig()
# Initialize shared components
self._store = vector_store
self._embedder = embedding_adapter
# Initialize agents
self._init_agents()
logger.info(
f"AgenticRAG initialized (model={self.config.model}, "
f"revision_attempts={self.config.max_revision_attempts})"
)
def _init_agents(self):
"""Initialize all agents with shared configuration."""
# Query Planner
self.planner = QueryPlannerAgent(
model=self.config.model,
base_url=self.config.base_url,
use_llm=self.config.enable_query_planning,
)
# Retriever
retriever_config = HybridSearchConfig(
dense_top_k=self.config.retrieval_top_k,
sparse_top_k=self.config.retrieval_top_k,
final_top_k=self.config.retrieval_top_k,
)
self.retriever = RetrieverAgent(
config=retriever_config,
vector_store=self._store,
embedding_adapter=self._embedder,
)
# Reranker
reranker_config = RerankerConfig(
model=self.config.model,
base_url=self.config.base_url,
top_k=self.config.final_top_k,
use_llm_rerank=self.config.enable_reranking,
min_relevance_score=0.1, # Lower threshold to allow more results
)
self.reranker = RerankerAgent(config=reranker_config)
# Synthesizer
synth_config = SynthesizerConfig(
model=self.config.model,
base_url=self.config.base_url,
confidence_threshold=self.config.min_confidence,
)
self.synthesizer = SynthesizerAgent(config=synth_config)
# Critic
critic_config = CriticConfig(
model=self.config.model,
base_url=self.config.base_url,
)
self.critic = CriticAgent(config=critic_config)
@property
def store(self) -> VectorStore:
"""Get vector store (lazy initialization)."""
if self._store is None:
self._store = get_vector_store()
return self._store
@property
def embedder(self) -> EmbeddingAdapter:
"""Get embedding adapter (lazy initialization)."""
if self._embedder is None:
self._embedder = get_embedding_adapter()
return self._embedder
def query(
self,
question: str,
filters: Optional[Dict[str, Any]] = None,
) -> RAGResponse:
"""
Process a query through the full RAG pipeline.
Args:
question: User's question
filters: Optional metadata filters for retrieval
Returns:
RAGResponse with answer and metadata
"""
# Initialize state
state = RAGState(query=question)
try:
# Stage 1: Query Planning
state = self._plan(state)
# Stage 2: Retrieval
state = self._retrieve(state, filters)
# Stage 3: Reranking
state = self._rerank(state)
# Stage 4: Synthesis
state = self._synthesize(state)
# Stage 5: Validation + Revision Loop
if self.config.enable_validation:
state = self._validate_and_revise(state)
# Build response
return self._build_response(state)
except Exception as e:
logger.error(f"RAG pipeline error: {e}")
state.errors.append(str(e))
return self._build_error_response(state, str(e))
def query_stream(
self,
question: str,
filters: Optional[Dict[str, Any]] = None,
) -> Generator[Tuple[PipelineStage, Any], None, None]:
"""
Process query with streaming updates.
Yields:
Tuple of (stage, stage_result)
"""
state = RAGState(query=question)
try:
# Planning
state = self._plan(state)
yield PipelineStage.PLANNING, state.query_plan
# Retrieval
state = self._retrieve(state, filters)
yield PipelineStage.RETRIEVAL, len(state.retrieved_chunks)
# Reranking
state = self._rerank(state)
yield PipelineStage.RERANKING, len(state.ranked_chunks)
# Synthesis
state = self._synthesize(state)
yield PipelineStage.SYNTHESIS, state.synthesis_result
# Validation
if self.config.enable_validation:
state = self._validate_and_revise(state)
yield PipelineStage.VALIDATION, state.critic_result
# Complete
response = self._build_response(state)
yield PipelineStage.COMPLETE, response
except Exception as e:
logger.error(f"Streaming error: {e}")
yield PipelineStage.COMPLETE, self._build_error_response(state, str(e))
def _plan(self, state: RAGState) -> RAGState:
"""Execute query planning stage."""
start = time.time()
state.stage = PipelineStage.PLANNING
if self.config.verbose:
logger.info(f"Planning query: {state.query}")
state.query_plan = self.planner.plan(state.query)
state.stage_times["planning"] = time.time() - start
if self.config.verbose:
logger.info(
f"Query plan: intent={state.query_plan.intent}, "
f"sub_queries={len(state.query_plan.sub_queries)}"
)
return state
def _retrieve(
self,
state: RAGState,
filters: Optional[Dict[str, Any]],
) -> RAGState:
"""Execute retrieval stage."""
start = time.time()
state.stage = PipelineStage.RETRIEVAL
if self.config.verbose:
logger.info("Retrieving relevant chunks...")
# Use hybrid retrieval with query plan
state.retrieved_chunks = self.retriever.retrieve(
query=state.query,
plan=state.query_plan,
top_k=self.config.retrieval_top_k,
filters=filters,
)
state.stage_times["retrieval"] = time.time() - start
if self.config.verbose:
logger.info(f"Retrieved {len(state.retrieved_chunks)} chunks")
return state
def _rerank(self, state: RAGState) -> RAGState:
"""Execute reranking stage."""
start = time.time()
state.stage = PipelineStage.RERANKING
if not state.retrieved_chunks:
state.ranked_chunks = []
return state
if self.config.verbose:
logger.info("Reranking results...")
state.ranked_chunks = self.reranker.rerank(
query=state.query,
results=state.retrieved_chunks,
top_k=self.config.final_top_k,
)
state.stage_times["reranking"] = time.time() - start
if self.config.verbose:
logger.info(f"Reranked to {len(state.ranked_chunks)} chunks")
return state
def _synthesize(self, state: RAGState) -> RAGState:
"""Execute synthesis stage."""
start = time.time()
state.stage = PipelineStage.SYNTHESIS
if self.config.verbose:
logger.info("Synthesizing answer...")
state.synthesis_result = self.synthesizer.synthesize(
query=state.query,
results=state.ranked_chunks,
plan=state.query_plan,
)
state.stage_times["synthesis"] = time.time() - start
if self.config.verbose:
logger.info(
f"Synthesized answer (confidence={state.synthesis_result.confidence:.2f})"
)
return state
def _validate_and_revise(self, state: RAGState) -> RAGState:
"""Execute validation and optional revision loop."""
start = time.time()
while state.revision_attempt <= self.config.max_revision_attempts:
state.stage = PipelineStage.VALIDATION
if self.config.verbose:
logger.info(f"Validating (attempt {state.revision_attempt + 1})...")
# Validate current synthesis
state.critic_result = self.critic.validate(
synthesis_result=state.synthesis_result,
sources=state.ranked_chunks,
)
if state.critic_result.is_valid:
if self.config.verbose:
logger.info("Validation passed!")
break
# Check if we should revise
if state.revision_attempt >= self.config.max_revision_attempts:
if self.config.verbose:
logger.warning("Max revision attempts reached")
break
# Attempt revision
state.stage = PipelineStage.REVISION
state.revision_attempt += 1
state.revision_history.append(state.synthesis_result)
if self.config.verbose:
logger.info(f"Revising answer (attempt {state.revision_attempt})...")
# Re-synthesize with critic feedback
state.synthesis_result = self._revise_synthesis(state)
state.stage_times["validation"] = time.time() - start
return state
def _revise_synthesis(self, state: RAGState) -> SynthesisResult:
"""Revise synthesis based on critic feedback."""
# Add revision hints to the synthesis prompt
# For now, just re-synthesize (a more advanced version would
# incorporate critic feedback into the prompt)
return self.synthesizer.synthesize(
query=state.query,
results=state.ranked_chunks,
plan=state.query_plan,
)
def _build_response(self, state: RAGState) -> RAGResponse:
"""Build final response from state."""
total_time = (time.time() - state.start_time) * 1000 # ms
synthesis = state.synthesis_result
if synthesis is None:
return self._build_error_response(state, "No synthesis result")
# Build query plan dict for response
query_plan_dict = None
if state.query_plan:
query_plan_dict = {
"intent": state.query_plan.intent.value,
"sub_queries": len(state.query_plan.sub_queries),
"expanded_terms": state.query_plan.expanded_terms[:5],
}
# Build validation dict
validation_dict = None
if state.critic_result:
validation_dict = {
"is_valid": state.critic_result.is_valid,
"confidence": state.critic_result.confidence,
"hallucination_score": state.critic_result.hallucination_score,
"citation_accuracy": state.critic_result.citation_accuracy,
"issues": len(state.critic_result.issues),
}
return RAGResponse(
answer=synthesis.answer,
citations=synthesis.citations,
confidence=synthesis.confidence,
query=state.query,
num_sources=synthesis.num_sources_used,
validated=state.critic_result.is_valid if state.critic_result else False,
revision_attempts=state.revision_attempt,
query_plan=query_plan_dict,
validation_details=validation_dict,
latency_ms=total_time,
)
def _build_error_response(
self,
state: RAGState,
error: str,
) -> RAGResponse:
"""Build error response."""
return RAGResponse(
answer=f"I encountered an error processing your query: {error}",
citations=[],
confidence=0.0,
query=state.query,
num_sources=0,
validated=False,
revision_attempts=state.revision_attempt,
latency_ms=(time.time() - state.start_time) * 1000,
)
def index_text(
self,
text: str,
document_id: str,
metadata: Optional[Dict[str, Any]] = None,
) -> int:
"""
Index text content into the vector store.
Args:
text: Text content to index
document_id: Unique document identifier
metadata: Optional metadata
Returns:
Number of chunks indexed
"""
# Simple chunking
chunk_size = 500
overlap = 50
chunks = []
embeddings = []
for i in range(0, len(text), chunk_size - overlap):
chunk_text = text[i:i + chunk_size]
if len(chunk_text.strip()) < 50:
continue
chunk_id = f"{document_id}_chunk_{len(chunks)}"
chunks.append({
"chunk_id": chunk_id,
"document_id": document_id,
"text": chunk_text,
"page": 0,
"chunk_type": "text",
"source_path": metadata.get("filename", "") if metadata else "",
})
# Generate embedding
embedding = self.embedder.embed_text(chunk_text)
embeddings.append(embedding)
if not chunks:
return 0
# Add to store
self.store.add_chunks(chunks, embeddings)
logger.info(f"Indexed {len(chunks)} chunks for document {document_id}")
return len(chunks)
def get_stats(self) -> Dict[str, Any]:
"""Get system statistics."""
return {
"total_chunks": self.store.count(),
"model": self.config.model,
"embedding_model": self.embedder.model_name,
"embedding_dimension": self.embedder.embedding_dimension,
}