File size: 12,898 Bytes
c75184e 4f811d2 c75184e 4f811d2 c75184e 6951cf0 c75184e bdc7a4a c75184e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 | """
LLM Service — Gemini (primary) + Groq (fallback) orchestrator.
Phase 2: Gemini Integration.
- When USE_GEMINI=true (default), queries go through GeminiService.
- When USE_GEMINI=false or GEMINI_API_KEY is missing, falls back to Groq.
- _get_llm: Groq LLM factory (preserved for fallback)
- MathAIEngine: Orchestrates retrieval, symbolic math, and LLM calls.
"""
import re
import time
import logging
from typing import Dict, Any, Optional, Tuple
from backend.src.config import settings, SYSTEM_TEMPLATE
from backend.src.math.symbolic_engine import SymbolicMathEngine
from backend.src.services.memory_service import MongoDBChatMemory
try:
from langchain_core.messages import HumanMessage, AIMessage, SystemMessage
except ImportError:
from langchain.schema import HumanMessage, AIMessage, SystemMessage
logger = logging.getLogger("math_assistant.llm")
_LLM_CACHE = {}
class GroqLLMWrapper:
"""Wrapper around native Groq client to mimic Langchain's invoke() and astream() methods."""
def __init__(self, api_key: str, model_name: str):
from groq import Groq, AsyncGroq
self.client = Groq(api_key=api_key)
self.async_client = AsyncGroq(api_key=api_key)
self.model_name = model_name
def invoke(self, messages):
from langchain_core.messages import AIMessage
groq_msgs = []
# If it's just a single string prompt
if isinstance(messages, str):
groq_msgs.append({"role": "user", "content": messages})
elif isinstance(messages, list):
for m in messages:
if isinstance(m, str):
groq_msgs.append({"role": "user", "content": m})
elif isinstance(m, dict):
# In case it's a raw dict
groq_msgs.append(m)
else:
role = "user" if getattr(m, "type", "user") == "human" else getattr(m, "type", "user")
content = m.content if isinstance(m.content, str) else str(m.content)
groq_msgs.append({"role": role, "content": content})
chat_completion = self.client.chat.completions.create(
messages=groq_msgs,
model=self.model_name,
temperature=0.1,
max_tokens=2048,
)
return AIMessage(content=chat_completion.choices[0].message.content)
async def astream(self, messages):
from langchain_core.messages import AIMessageChunk
groq_msgs = []
if isinstance(messages, str):
groq_msgs.append({"role": "user", "content": messages})
elif isinstance(messages, list):
for m in messages:
if isinstance(m, str):
groq_msgs.append({"role": "user", "content": m})
elif isinstance(m, dict):
groq_msgs.append(m)
else:
role = "user" if getattr(m, "type", "user") == "human" else getattr(m, "type", "user")
content = m.content if isinstance(m.content, str) else str(m.content)
groq_msgs.append({"role": role, "content": content})
stream = await self.async_client.chat.completions.create(
messages=groq_msgs,
model=self.model_name,
temperature=0.1,
max_tokens=2048,
stream=True
)
async for chunk in stream:
content = chunk.choices[0].delta.content
if content is not None:
yield AIMessageChunk(content=content)
def _get_llm(model=None):
"""Get or create a cached Groq LLM instance for a given model name."""
key = model or settings.groq_model_fallbacks[0]
if key not in _LLM_CACHE:
api_key = settings.groq_api_key
if not api_key:
raise ValueError("GROQ_API_KEY not set. Add it to your .env file.")
logger.info(f"Initializing Native Groq LLM: {key}")
_LLM_CACHE[key] = GroqLLMWrapper(api_key=api_key, model_name=key)
return _LLM_CACHE[key]
class MathAIEngine:
"""Orchestrates RAG retrieval, symbolic math, and LLM calls for math tutoring."""
def __init__(self, vector_store=None, session_id: str = "default"):
try:
self.llm = self._init_llm() if not settings.use_gemini else None
except Exception as e:
logger.warning(f"Failed to initialize Groq LLM: {e}")
self.llm = None
self.vector_store = vector_store
self.memory = MongoDBChatMemory(session_id=session_id)
self.symbolic = SymbolicMathEngine()
self.session_id = session_id
def _init_llm(self):
"""Initialize the fallback Groq LLM model."""
return _get_llm(settings.groq_model_fallbacks[0])
def _retrieve_context(self, query: str) -> Tuple[list, str]:
"""Retrieve relevant documents from the vector store."""
if not self.vector_store or not self.vector_store.is_ready():
return [], "No knowledge base available. Using general mathematical knowledge."
docs = self.vector_store.similarity_search(query, k=3)
if not docs:
return [], "No specific context found."
parts = [f"[Reference {i+1} - {d.metadata.get('topic','math')}]\n{d.page_content}"
for i, d in enumerate(docs)]
return docs, "\n\n---\n\n".join(parts)
def _symbolic_hint(self, query: str) -> Optional[str]:
"""Try to get a symbolic verification hint for the query."""
ql = query.lower()
for pattern, action in [
(r"(?:differentiate|derivative of|d/dx)\s+(.+?)(?:\s+with respect|\s*$)", "diff"),
(r"(?:integrate|integral of)\s+(.+?)(?:\s+with respect|\s+dx|\s*$)", "int"),
(r"solve\s+(.+?)\s+(?:for|=)", "solve"),
]:
m = re.search(pattern, ql)
if m:
expr = m.group(1).strip()
result = (self.symbolic.differentiate(expr) if action == "diff"
else self.symbolic.integrate(expr) if action == "int"
else self.symbolic.solve_equation(expr))
if result:
return f"[Symbolic verification: {result}]"
return None
def query(self, user_input: str) -> Dict[str, Any]:
"""Process a user math query through the full RAG pipeline.
Flow: symbolic hint → vector retrieval → chat history →
Gemini (primary) or Groq (fallback) → store response.
"""
hint = self._symbolic_hint(user_input)
source_docs, context = self._retrieve_context(user_input)
chat_history = self.memory.get_langchain_messages(limit=4)
# Store human message BEFORE LLM call so history order is correct
self.memory.add_message("human", user_input)
# ── Try Gemini first (Phase 2) ─────────────────────────────────
if settings.use_gemini and settings.gemini_api_key:
try:
from backend.src.services.gemini_service import GeminiService
gemini = GeminiService()
answer = gemini.query(user_input, context=context, chat_history=chat_history)
if answer and not answer.startswith("⚠️"):
self.memory.add_message("assistant", answer)
sources = [{"topic": d.metadata.get("topic", "unknown"),
"source": d.metadata.get("source", "kb"),
"difficulty": d.metadata.get("difficulty", "unknown")}
for d in source_docs]
return {"answer": answer, "sources": sources, "symbolic_hint": hint,
"session_id": self.session_id, "context_docs": len(source_docs)}
except Exception as e:
logger.warning(f"Gemini failed, falling back to Groq: {e}")
# ── Groq Fallback ──────────────────────────────────────────────
# Build messages list manually — NEVER pass math content through
# LangChain format_messages(), because curly braces in math break templates
system_text = SYSTEM_TEMPLATE.replace("{context}", context)
llm_messages = [SystemMessage(content=system_text)]
for msg in chat_history:
llm_messages.append(msg)
llm_messages.append(HumanMessage(content=user_input))
# Auto-retry with model fallback
answer = None
last_error = None
used_models = []
for model_name in settings.groq_model_fallbacks:
if model_name in used_models:
continue
used_models.append(model_name)
llm = _get_llm(model_name)
for _attempt in range(2):
try:
raw = llm.invoke(llm_messages).content
if raw and not any(p in raw.lower() for p in
["rate limit", "too many requests", "service unavailable"]):
answer = raw
if model_name != settings.groq_model_fallbacks[0]:
answer = f"*(Using fallback model: {model_name})*\n\n" + answer
break
else:
raise Exception(raw or "Empty response")
except Exception as e:
last_error = e
err = str(e).lower()
full_err = str(e)
logger.warning(f"Model {model_name} attempt {_attempt+1} failed: {e}")
if "per day" in full_err or "tokens per day" in full_err:
logger.info(f"Daily limit on {model_name}, trying next model...")
break
elif "429" in full_err or "rate_limit" in err:
time.sleep(2 ** _attempt)
continue
elif "timeout" in err or "connect" in err or "503" in err:
time.sleep(1)
continue
else:
break
if answer:
break
if answer is None:
full_err = str(last_error)
err = full_err.lower()
logger.error(f"LLM failed: [{type(last_error).__name__}] {full_err}")
if "401" in full_err or "invalid_api_key" in err:
answer = "⚠️ Invalid API key. Check GROQ_API_KEY in your .env file."
elif "429" in full_err or "rate_limit_exceeded" in err:
retry_match = re.search(r'try again in (.+?)\.', full_err)
retry_info = f" Groq says: try again in **{retry_match.group(1)}**." if retry_match else ""
if "per day" in full_err or "tokens per day" in full_err or "TPD" in full_err:
answer = (f"⚠️ **Daily token limit reached** (Groq free tier: 100,000 tokens/day).{retry_info}\n\n"
f"To keep using the app now, change your `.env`:\n```\nLLM_MODEL=llama3-8b-8192\n```\n"
f"Then restart Streamlit. The 8B model has a separate 500k/day quota.")
else:
answer = f"⚠️ **Rate limit hit** (too many requests per minute).{retry_info} Wait 20–30 seconds and try again."
elif "context_length" in err or ("context" in err and "length" in err):
answer = "⚠️ Question + context too long. Try a shorter question."
elif "connect" in err or "connection" in err:
answer = "⚠️ Cannot reach Groq API. Check your internet connection."
elif "timeout" in err:
answer = "⚠️ Request timed out. Try again."
else:
answer = f"⚠️ Error ({type(last_error).__name__}): {full_err}"
self.memory.add_message("assistant", answer)
sources = [{"topic": d.metadata.get("topic", "unknown"),
"source": d.metadata.get("source", "kb"),
"difficulty": d.metadata.get("difficulty", "unknown")}
for d in source_docs]
return {"answer": answer, "sources": sources, "symbolic_hint": hint,
"session_id": self.session_id, "context_docs": len(source_docs)}
def clear_memory(self):
"""Clear chat history for this session."""
self.memory.clear_history()
def get_history(self):
"""Get full chat history for this session."""
return self.memory.get_history(limit=50)
|