import os from dotenv import load_dotenv from langsmith import traceable # CRITICAL: For observability from langchain_google_genai import ChatGoogleGenerativeAI, GoogleGenerativeAIEmbeddings from langchain_chroma import Chroma # Importing your logic from other files from src.retrieval.reranker import get_reranked_docs from src.agents.final_agent import generate_answer load_dotenv() # --- 1. SET UP MODELS & CACHE --- embeddings = GoogleGenerativeAIEmbeddings(model="models/gemini-embedding-2-preview") cache_db = Chroma( collection_name="llm_semantic_cache", persist_directory="./chroma_db", embedding_function=embeddings ) # --- 2. TRACEABLE SUB-COMPONENTS --- # We wrap these so they show up as separate bars in the LangSmith UI @traceable(name="Document Retrieval & Rerank", run_type="retriever") def traced_retrieval(query): return get_reranked_docs(query) @traceable(name="LLM Answer Generation", run_type="llm") def traced_generation(query, docs): return generate_answer(query, docs) @traceable(name="Faithfulness Evaluation", run_type="chain") def traced_evaluation(llm, context, answer): eval_prompt = f"Context: {context}\nAnswer: {answer}\nIs this answer supported? YES/NO" response = llm.invoke(eval_prompt) # --- GEMINI 3 CONTENT EXTRACTION --- # The response.content is now a list like: [{'type': 'text', 'text': 'YES', 'extras': {...}}] if isinstance(response.content, list): # Extract the 'text' key from the first block that has it content = next((block['text'] for block in response.content if 'text' in block), "") else: # Fallback for older models/strings content = str(response.content) return content.strip().upper() # Force to uppercase for reliable YES/NO check # --- 3. MAIN TRACED PIPELINE --- @traceable(name="Nutrition RAG Pipeline", run_type="chain") def run_with_cache_and_eval(query): # Initialize defaults to prevent "got 0" errors res, tokens, verd, ctx = "No response", 0, "FAIL", "No context" try: llm = ChatGoogleGenerativeAI(model="gemini-3-flash-preview", temperature=0) # 1. Check Cache cache_results = cache_db.similarity_search_with_relevance_scores(query, k=1) if cache_results and cache_results[0][1] > 0.92: metadata = cache_results[0][0].metadata return metadata.get("response"), 0, "YES (Cached)", metadata.get("context", "No cached context") # 2. Execution Flow final_docs = traced_retrieval(query) if not final_docs: return "No documents found", 0, "NO_DATA", "Context empty" ctx = "\n\n".join([d.page_content for d in final_docs]) res = traced_generation(query, final_docs) tokens = llm.get_num_tokens(ctx + query) # 3. Audit Flow verd = traced_evaluation(llm, ctx, res) # 4. Persistence cache_db.add_texts( texts=[query], metadatas=[{"response": res, "context": ctx}] ) return res, tokens, verd, ctx except Exception as e: import traceback error_details = traceback.format_exc() print(f"--- INTERNAL PIPELINE CRASH ---") print(error_details) print(f"-------------------------------") # Return the SPECIFIC traceback so the Judge (and you) can see it return f"CRASH in pipeline: {str(e)}", 0, "ERROR", "No context" if __name__ == "__main__": q = "give me recipes with > 20g protein and < 500 calories" ans, tokens, verdict, context = run_with_cache_and_eval(q) print("\n" + "="*50) print("ARCHITECT'S OPS REPORT (Traced via LangSmith)") print("="*50) print(f"VERDICT: {verdict} | TOKENS: {tokens}") print(f"RESULT: {ans[:100]}...") # Printing first 100 chars print(f"CONTEXT: {context[:100]}...") # Printing first 100 chars print("="*50)