petProject / qa_backend.py
nesanchezo's picture
Update qa_backend.py
e5fa210 verified
import logging
import streamlit as st
from typing import Dict, List, Any
import pickle
import chromadb
from chromadb.config import Settings
from openai import OpenAI
import numpy as np
from nltk.tokenize import word_tokenize
from dotenv import load_dotenv
import os
from langsmith import traceable
from langsmith import Client
from langsmith.wrappers import wrap_openai
# Configure logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
# Load environment variables
#load_dotenv()
# Obtener las claves de los secrets de Hugging Face
#OpenAI.api_key = st.secrets["OPENAI_API_KEY"].strip()
#os.environ["LANGCHAIN_API_KEY"] = st.secrets["LANGCHAIN_API_KEY"]
#os.environ["LANGCHAIN_TRACING_V2"] = st.secrets["LANGCHAIN_TRACING_V2"]
# Initialize LangSmith client
langsmith_client = Client()
# Wrap OpenAI client with LangSmith
openai = wrap_openai(OpenAI())
def detect_language(text: str) -> str:
"""
Simple language detection for English/Spanish based on common words.
Args:
text: Input text to detect language
Returns:
str: 'es' for Spanish, 'en' for English
"""
# Common Spanish words/characters
spanish_indicators = {'qué', 'cuál', 'cómo', 'dónde', 'por', 'para', 'perro', 'comida',
'mejor', 'precio', 'barato', 'caro', 'cachorro', 'adulto'}
# Convert to lowercase for comparison
text_lower = text.lower()
# Count Spanish indicators
spanish_count = sum(1 for word in spanish_indicators if word in text_lower)
# If we find Spanish indicators, classify as Spanish, otherwise default to English
return 'es' if spanish_count > 0 else 'en'
class DogFoodQASystem:
def __init__(self):
"""Initialize the QA system with vector stores and models."""
self.load_stores()
def load_stores(self) -> None:
"""Load BM25 and ChromaDB stores."""
with open('bm25_index.pkl', 'rb') as f:
self.bm25_data = pickle.load(f)
self.chroma_client = chromadb.PersistentClient(path="chroma_db")
self.collection = self.chroma_client.get_collection("dog_food_descriptions")
@traceable(run_type="chain")
def hybrid_search(self, query: str, top_k: int = 5) -> List[Dict[str, Any]]:
"""
Hybrid search that gets top_k results from each source and combines unique results.
"""
logging.info(f"\n{'='*50}\nStarting hybrid search for query: {query}")
# BM25 search - get top_k results
tokenized_query = word_tokenize(query.lower())
bm25_scores = self.bm25_data['model'].get_scores(tokenized_query)
bm25_indices = np.argsort(bm25_scores)[::-1][:top_k] # Get top_k results
bm25_results = [
{
'score': float(bm25_scores[idx]),
'text': self.bm25_data['chunks'][idx],
'metadata': self.bm25_data['metadata'][idx],
'source': 'BM25'
}
for idx in bm25_indices
]
logging.info(f"Retrieved {len(bm25_results)} results from BM25")
# Vector search - get top_k results
try:
embedding_response = openai.embeddings.create(
model="text-embedding-ada-002",
input=query
)
query_embedding = embedding_response.data[0].embedding
chroma_results = self.collection.query(
query_embeddings=[query_embedding],
n_results=top_k, # Get top_k results
include=["documents", "metadatas", "distances"]
)
processed_vector_results = [
{
'score': float(1 - distance),
'text': doc,
'metadata': meta,
'source': 'Vector'
}
for doc, meta, distance in zip(
chroma_results['documents'][0],
chroma_results['metadatas'][0],
chroma_results['distances'][0]
)
]
logging.info(f"Retrieved {len(processed_vector_results)} results from Vector search")
except Exception as e:
logging.error(f"Error in vector search: {str(e)}", exc_info=True)
processed_vector_results = []
# Combine results
all_results = self._smart_combine_results(bm25_results, processed_vector_results, query)
return all_results
def _smart_combine_results(self, bm25_results: List[Dict], vector_results: List[Dict], query: str) -> List[Dict]:
"""
Combine results from both sources, tracking duplicates and sources.
"""
logging.info("\nCombining search results...")
# Initialize combined results with tracking
combined_dict = {} # Use text as key to track duplicates
# Process vector results
for result in vector_results:
text = result['text']
if text not in combined_dict:
result['sources'] = ['Vector']
result['original_scores'] = {'Vector': result['score']}
combined_dict[text] = result
logging.info(f"Added Vector result (score: {result['score']:.4f})")
else:
combined_dict[text]['sources'].append('Vector')
combined_dict[text]['original_scores']['Vector'] = result['score']
logging.info(f"Marked existing result as found by Vector (score: {result['score']:.4f})")
# Process BM25 results
for result in bm25_results:
text = result['text']
if text not in combined_dict:
result['sources'] = ['BM25']
result['original_scores'] = {'BM25': result['score']}
combined_dict[text] = result
logging.info(f"Added BM25 result (score: {result['score']:.4f})")
else:
combined_dict[text]['sources'].append('BM25')
combined_dict[text]['original_scores']['BM25'] = result['score']
logging.info(f"Marked existing result as found by BM25 (score: {result['score']:.4f})")
# Convert to list
all_results = list(combined_dict.values())
# Calculate statistics
total_results = len(all_results)
duplicates = sum(1 for r in all_results if len(r['sources']) > 1)
vector_only = sum(1 for r in all_results if r['sources'] == ['Vector'])
bm25_only = sum(1 for r in all_results if r['sources'] == ['BM25'])
logging.info(f"\nResults Statistics:")
logging.info(f"- Total unique results: {total_results}")
logging.info(f"- Duplicates (found by both): {duplicates}")
logging.info(f"- Vector only: {vector_only}")
logging.info(f"- BM25 only: {bm25_only}")
return all_results
def _adjust_score_with_metadata(self, result: Dict, query: str) -> float:
"""Adjust search score based on metadata relevance."""
base_score = result['score']
metadata = result['metadata']
# Initialize boost factors
boost = 1.0
# Boost based on reviews (social proof)
if metadata.get('reviews', 0) > 20:
boost *= 1.2
# Boost based on price range mentions
query_lower = query.lower()
if ('affordable' in query_lower or 'barato' in query_lower) and metadata.get('price', 0) < 50:
boost *= 1.3
elif ('premium' in query_lower or 'premium' in query_lower) and metadata.get('price', 0) > 100:
boost *= 1.3
# Boost based on specific dog type matches
dog_types = ['puppy', 'adult', 'senior', 'cachorro', 'adulto']
for dog_type in dog_types:
if dog_type in query_lower and dog_type in metadata.get('dog_type', '').lower():
boost *= 1.25
break
return base_score * boost
@traceable(run_type="chain")
def generate_answer(self, query: str, search_results: List[Dict]) -> str:
"""Generate a natural language answer based on search results."""
# Detect query language
query_lang = detect_language(query)
# Prepare context from search results
context = self._prepare_context(search_results)
# Create prompt based on language
system_prompt = {
'es': """Eres un experto en nutrición canina. Responde a la pregunta utilizando solo el contexto proporcionado.
Si no puedes responder con el contexto dado, indícalo. Incluye información sobre precios y características
específicas de los productos cuando sea relevante.""",
'en': """You are a dog nutrition expert. Answer the question using only the provided context.
If you cannot answer from the given context, say so. Include pricing and specific product
features when relevant."""
}.get(query_lang, 'en')
response = openai.chat.completions.create(
model="gpt-4o-mini",
messages=[
{"role": "system", "content": system_prompt},
{"role": "user", "content": f"Context:\n{context}\n\nQuestion: {query}"}
],
temperature=0.7,
max_tokens=300
)
return response.choices[0].message.content.strip()
def _prepare_context(self, search_results: List[Dict]) -> str:
"""Prepare search results as context for the LLM."""
context_parts = []
for result in search_results:
metadata = result['metadata']
context_parts.append(
f"Product: {metadata['product_name']}\n"
f"Brand: {metadata['brand']}\n"
f"Price: ${metadata['price']}\n"
f"Weight: {metadata['weight']}kg\n"
f"Dog Type: {metadata['dog_type']}\n"
f"Description: {result['text']}\n"
)
return "\n---\n".join(context_parts)
@traceable(run_type="chain")
def process_query(self, query: str) -> Dict[str, Any]:
"""Process a user query and return both search results and answer."""
search_results = self.hybrid_search(query)
answer = self.generate_answer(query, search_results)
return {
"answer": answer,
"search_results": search_results,
"language": detect_language(query)
}
def diagnose_vector_store(self):
"""Diagnose the vector store setup."""
try:
logging.info("\nDiagnosing Vector Store:")
collection_info = self.collection.get()
# Basic collection info
doc_count = len(collection_info['ids'])
logging.info(f"Collection name: {self.collection.name}")
logging.info(f"Number of documents: {doc_count}")
# Sample query test
if doc_count > 0:
test_query = "test query for diagnosis"
test_embedding = openai.embeddings.create(
model="text-embedding-ada-002",
input=test_query
).data[0].embedding
test_results = self.collection.query(
query_embeddings=[test_embedding],
n_results=1
)
if len(test_results['ids'][0]) > 0:
logging.info("✅ Vector store test query successful")
return True
else:
logging.error("❌ Vector store returned no results for test query")
return False
else:
logging.error("❌ Vector store is empty")
return False
except Exception as e:
logging.error(f"❌ Error accessing vector store: {str(e)}")
return False