dev-models's picture
Initial commit
e97c8d1
import json
import os
import numpy as np
import torch
from typing import List, Dict, Any, Optional
from tqdm import tqdm
from pymongo import ReplaceOne
from rank_bm25 import BM25Okapi
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from config import VECTOR_INDEX_NAME
from .database import get_mongo_client, get_mongo_collection
from .models import get_clip_model, get_llm, get_groq_client
from dotenv import load_dotenv
import time
load_dotenv()
import os
class RAGEngine:
"""
Unified RAG engine refactored from search.py.
"""
def __init__(self, use_hybrid: bool = True, force_clean: bool = False):
self.use_hybrid = use_hybrid
self.clip_model = get_clip_model()
self.collection = get_mongo_collection()
self.llm = get_llm()
self.groq_client = get_groq_client()
if force_clean:
self.collection.delete_many({})
self._setup_vector_index()
self.bm25_index = None
self.bm25_doc_map = {}
if self.collection.count_documents({}) > 0:
self._rebuild_bm25_index()
def _setup_vector_index(self):
"""
Attempts to create a vector search index if using MongoDB Atlas.
Includes robust dimension checking and error handling.
"""
# 1. Determine Dimensions safely
try:
dims = self.clip_model.get_sentence_embedding_dimension()
if dims is None or not isinstance(dims, int):
raise ValueError("Model returned invalid dimensions")
except Exception:
print("Auto-dim failed, probing model...")
test_vec = self.clip_model.encode("test")
dims = len(test_vec)
print(f"Vector Dimensions: {dims}")
# 2. Define Index Model
index_model = {
"definition": {
"fields": [
{
"type": "vector",
"path": "embedding",
"numDimensions": int(dims), # Ensure strict integer
"similarity": "cosine"
},
{
"type": "filter",
"path": "metadata.type"
}
]
},
"name": VECTOR_INDEX_NAME,
"type": "vectorSearch"
}
# 3. Create Index
try:
# Check if index already exists
indexes = list(self.collection.list_search_indexes())
index_names = [idx.get("name") for idx in indexes]
if VECTOR_INDEX_NAME not in index_names:
print(f"Creating Atlas Vector Search Index '{VECTOR_INDEX_NAME}'...")
self.collection.create_search_index(model=index_model)
print("Index creation initiated. Please wait 1-2 minutes for Atlas to build it.")
print("You can check progress in Atlas UI -> Database -> Search -> Vector Search")
else:
print(f"Index '{VECTOR_INDEX_NAME}' already exists.")
except Exception as e:
print(f"\nAutomatic Index Creation Failed: {e}")
print("This is common on Free Tier (M0) or due to permissions.")
print("PLEASE CREATE MANUALLY IN ATLAS UI (See JSON below)\n")
print(json.dumps(index_model["definition"], indent=2))
except Exception as e:
print(f"Unexpected error checking/creating index: {e}")
def _rebuild_bm25_index(self):
cursor = self.collection.find(
{"metadata.type": {"$in": ["text", "table", "list", "header", "code"]}},
{"content": 1, "_id": 1}
)
text_docs = []
self.bm25_doc_map = {}
for idx, doc in enumerate(cursor):
content = doc.get("content", "")
if content:
text_docs.append(content.lower().split())
self.bm25_doc_map[idx] = str(doc["_id"])
if text_docs:
self.bm25_index = BM25Okapi(text_docs)
def _encode_content(self, content: Any, content_type: str) -> np.ndarray:
if content_type == "image":
# Assuming content is base64
from PIL import Image
from io import BytesIO
import base64
try:
img = Image.open(BytesIO(base64.b64decode(content))).convert("RGB")
return self.clip_model.encode(img, normalize_embeddings=True)
except: return None
return self.clip_model.encode(content, normalize_embeddings=True)
def ingest_data(self, data: Dict[str, Any]):
"""Ingests processed document data."""
operations = []
for chunk in data.get("chunks", []):
embedding = self._encode_content(chunk["text"], "text")
if embedding is None: continue
doc = {
"_id": chunk["chunk_id"],
"content": chunk["text"],
"embedding": embedding.tolist(),
"metadata": {
**chunk["metadata"],
"type": chunk.get("type", "text")
}
}
operations.append(ReplaceOne({"_id": doc["_id"]}, doc, upsert=True))
for img in data.get("images", []):
embedding = self._encode_content(img["image_base64"], "image")
if embedding is None: continue
doc = {
"_id": img["image_id"],
"content": img.get("description", ""),
"embedding": embedding.tolist(),
"metadata": {
"page": str(img.get("page_number", 0)),
"header": str(img.get("section_header", "")),
"type": "image",
"description": img.get("description", ""),
"image_base64": img["image_base64"]
}
}
operations.append(ReplaceOne({"_id": doc["_id"]}, doc, upsert=True))
if operations:
for i in range(0, len(operations), 100):
self.collection.bulk_write(operations[i:i+100])
self._rebuild_bm25_index()
def hybrid_search(self, query: str, top_k: int = 5, alpha: float = 0.5) -> List[Dict]:
query_embedding = self._encode_content(query, "text")
dense_results = []
try:
pipeline = [
{"$vectorSearch": {
"index": VECTOR_INDEX_NAME,
"path": "embedding",
"queryVector": query_embedding.tolist(),
"numCandidates": top_k * 10,
"limit": top_k * 2
}},
{"$project": {"content": 1, "metadata": 1, "score": {"$meta": "vectorSearchScore"}}}
]
dense_results = list(self.collection.aggregate(pipeline))
except: pass
dense_scores = {str(r["_id"]): {"score": r.get("score", 0), "doc": r} for r in dense_results}
sparse_scores = {}
if self.bm25_index:
scores = self.bm25_index.get_scores(query.lower().split())
max_s = max(scores) if len(scores) > 0 and max(scores) > 0 else 1.0
for i in np.argsort(scores)[::-1][:top_k*2]:
if scores[i] > 0:
sparse_scores[self.bm25_doc_map[i]] = scores[i] / max_s
combined = []
all_ids = set(dense_scores.keys()) | set(sparse_scores.keys())
for did in all_ids:
d_s = dense_scores.get(did, {}).get("score", 0)
s_s = sparse_scores.get(did, 0)
score = (alpha * d_s) + ((1-alpha) * s_s)
doc = dense_scores.get(did, {}).get("doc") or self.collection.find_one({"_id": did})
if doc:
combined.append({**doc, "score": score})
combined.sort(key=lambda x: x["score"], reverse=True)
return combined[:top_k]
def answer_question(self, question: str, top_k: int = 5) -> str:
results = self.hybrid_search(question, top_k=top_k)
if not results: return "No relevant info found."
context = ""
for i, res in enumerate(results, 1):
m = res["metadata"]
context += f"\n[Src {i} | Page {m.get('page_number','?')}] {res['content']}"
prompt = f"Context:\n{context}\n\nQuestion: {question}\n\nAnswer strictly based on context:"
try:
chain = ChatPromptTemplate.from_template("{p}") | self.llm | StrOutputParser()
# return chain.invoke({"p": prompt})
for msg in chain.stream({"p": prompt}):
if hasattr(msg, "content"):
time.sleep(0.01)
yield msg.content
else:
time.sleep(0.01)
yield str(msg)
except Exception as e: return f"Error: {e}"
def search_images(self, query: str, top_k: int = 3, min_score: float = 0.5) -> List[Dict]:
query_embedding = self._encode_content(f"{query}", "text")
try:
pipeline = [
{"$vectorSearch": {
"index": VECTOR_INDEX_NAME, "path": "embedding",
"queryVector": query_embedding.tolist(), "numCandidates": top_k*10, "limit": top_k*2,
"filter": {"metadata.type": "image"}
}},
{"$project": {"content": 1, "metadata": 1, "score": {"$meta": "vectorSearchScore"}}}
]
results = list(self.collection.aggregate(pipeline))
return [{"description": r["content"], "image_base64": r["metadata"].get("image_base64"), "score": r["score"]}
for r in results if r["score"] >= min_score][:top_k]
except Exception as e:
print("*********error", str(e))
return []
# def generate_suggested_questions(self, num_questions: int = 5) -> List[str]:
# # Simple metadata-based generation or just a fixed list for now
# return ["What is the main topic?", "Explain the diagrams.", "Summarize the results."]
def generate_suggested_questions(self, num_questions: int = 4) -> List[str]:
"""Token-efficient question generation using metadata."""
print("\nGenerating suggested questions (Efficient Mode)...")
try:
# 1. Fetch metadata ONLY (projection excludes embedding and content)
cursor = self.collection.find(
{},
{"metadata": 1, "_id": 0}
).limit(100)
metadatas = [doc.get('metadata', {}) for doc in cursor]
if not metadatas:
return ["What is this document about?"]
# 2. Extract High-Level Structure
headers = set()
image_descriptions = []
import random
random.shuffle(metadatas)
for meta in metadatas:
if 'header' in meta and len(headers) < 8:
h = str(meta['header']).strip()
if h and h.lower() != "unknown" and len(h) > 5:
headers.add(h)
if meta.get('type') == 'image' and len(image_descriptions) < 2:
desc = meta.get('description', '')
if len(desc) > 20:
image_descriptions.append(desc[:100] + "...")
# 3. Construct Prompt
context_str = "Document Sections:\n" + "\n".join([f"- {h}" for h in headers])
if image_descriptions:
context_str += "\n\nVisual Content involves:\n" + "\n".join([f"- {d}" for d in image_descriptions])
# 4. Prompt LLM
prompt = f"""Generate {num_questions} short, interesting questions about a document with these sections and visuals:
{context_str}
Output ONLY the {num_questions} questions, one per line. No numbering."""
prompt_tmpl = ChatPromptTemplate.from_messages([
("system", "You are a helpful assistant."),
("user", "{prompt}")
])
chain = prompt_tmpl | self.llm | StrOutputParser()
response = chain.invoke({"prompt": prompt})
questions = [q.strip().lstrip('-1234567890. ') for q in response.split('\n') if q.strip()]
return questions[:num_questions]
except Exception as e:
print(f"Error generating questions: {e}")