import os from dotenv import load_dotenv from typing import List from pydantic import BaseModel, Field from langchain_core.output_parsers import PydanticOutputParser from langchain_google_genai import ChatGoogleGenerativeAI from langchain_core.prompts import ChatPromptTemplate # Import your optimized logic from previous pieces from src.retrieval.reranker import get_reranked_docs load_dotenv() # --- 1. Define the Data Contract --- class RecipeResponse(BaseModel): """Schema to force-strip metadata and signatures.""" answer: str = Field(description="The clean text answer containing recipes and protein counts.") def generate_answer(query, context_docs): # Initialize Gemini 3 Flash llm = ChatGoogleGenerativeAI(model="gemini-3-flash-preview", temperature=0) # --- 2. Initialize the Parser --- parser = PydanticOutputParser(pydantic_object=RecipeResponse) # 3. Enhanced Template with Format Instructions template = """ You are a professional Nutrition Assistant. Use the provided context to answer the question. {format_instructions} Context: {context} Question: {question} Instructions: - Only use the provided context. - Provide the Recipe Name and the specific Protein amount. """ prompt = ChatPromptTemplate.from_template( template, partial_variables={"format_instructions": parser.get_format_instructions()} ) # Prepare context context_text = "\n\n".join([doc.page_content for doc in context_docs]) # --- 4. The Clean Chain --- # The parser at the end of the chain is what wipes out the 'extras' dictionary chain = prompt | llm | parser response = chain.invoke({"context": context_text, "question": query}) # Return ONLY the string from our Pydantic object return response.answer if __name__ == "__main__": query = "give me protein recipes which can give me 25g protein" print("Agent is processing...") final_docs = get_reranked_docs(query) if final_docs: result = generate_answer(query, final_docs) print("\n--- AGENT ANSWER ---") print(result) # Now prints ONLY the validated text else: print("No relevant documents found.")