TickerNote-python-api / ragService.py
DevilxHacker
deploy TickerNote-python-api through spaces
6656c48
Raw
History Blame Contribute Delete
13.2 kB
import os
import time
import logging
from typing import List, Dict, Optional
from datetime import datetime
from pymongo import MongoClient, UpdateOne
from pymongo.operations import SearchIndexModel
from sentence_transformers import SentenceTransformer
import google.generativeai as genai
from dotenv import load_dotenv
load_dotenv()
logger = logging.getLogger(__name__)
# Config
MONGO_URI = os.getenv("MONGO_URI")
GEMINI_API_KEY = os.getenv("GEMINI_API_KEY")
DB_NAME = os.getenv("MONGO_DB_NAME", "ragdb")
COLLECTION_NAME = "rag_chunks"
VECTOR_INDEX = "vector_index"
EMBEDDING_DIM = 384
# retrival
TOP_K = 1
MIN_SCORE = 0.50
MAX_CHUNK_CHARS = 3000
# Context / prompt
MAX_CONTEXT_CHARS = 1200
CHUNK_TRUNCATE = 1000
MAX_HISTORY_TURNS = 1
HISTORY_MSG_CAP = 150
# Generation
MAX_OUTPUT_TOKENS = 256
TEMPERATURE = 0.2
# Retry
MAX_RETRIES = 3
RETRY_BASE_WAIT = 1.0
_embedder: Optional[SentenceTransformer] = None
_mongo_client: Optional[MongoClient] = None
_collection = None
def get_embedder() -> SentenceTransformer:
global _embedder
if _embedder is None:
logger.info("Loading sentence-transformer model...")
_embedder = SentenceTransformer("all-MiniLM-L6-v2")
logger.info("Model loaded.")
return _embedder
def get_collection():
global _mongo_client, _collection
if _collection is None:
_mongo_client = MongoClient(MONGO_URI)
db = _mongo_client[DB_NAME]
_collection = db[COLLECTION_NAME]
_ensure_vector_index(_collection)
return _collection
# ── Vector Index ──────────────────────────────────────────────────────────────
def _ensure_vector_index(col):
try:
db = col.database
if COLLECTION_NAME not in db.list_collection_names():
db.create_collection(COLLECTION_NAME)
logger.info(f"Created {COLLECTION_NAME} collection.")
existing = list(col.list_search_indexes())
names = [idx.get("name") for idx in existing]
if VECTOR_INDEX not in names:
logger.info("Creating Atlas Vector Search index...")
index_def = SearchIndexModel(
definition={
"fields": [
{
"type": "vector",
"path": "embedding",
"numDimensions": EMBEDDING_DIM,
"similarity": "cosine",
},
{
"type": "filter",
"path": "document_id",
},
]
},
name=VECTOR_INDEX,
type="vectorSearch",
)
col.create_search_index(index_def)
logger.info("Vector index created. It may take ~1 min to become active.")
else:
logger.info("Vector index already exists.")
except Exception as e:
logger.warning(
f"Could not verify/create vector index: {e}. "
"Ensure your Atlas cluster supports Vector Search."
)
# Embedding
def embed_texts(texts: List[str]) -> List[List[float]]:
model = get_embedder()
vecs = model.encode(
texts, batch_size=32, show_progress_bar=False, normalize_embeddings=True
)
return vecs.tolist()
def embed_query(text: str) -> List[float]:
return embed_texts([text])[0]
#Ingest
def ingest_chunks(chunks: List[Dict], document_id: str) -> int:
"""
Embed each chunk and upsert into MongoDB Atlas.
chunks : list of dicts from reportCleaning.process_pdf_to_chunks()
Expected keys: chunk_id, text, section, chunk_index,
total_chunks, pages, token_count, metadata
document_id : unique identifier (e.g. file._id from MongoDB)
Returns number of chunks stored.
"""
col = get_collection()
texts = [c["text"] for c in chunks]
logger.info(f"Embedding {len(texts)} chunks for document '{document_id}'...")
embeddings = embed_texts(texts)
logger.info("Embeddings done.")
ops = []
for chunk, emb in zip(chunks, embeddings):
doc = {
"document_id": document_id,
"chunk_id": chunk["chunk_id"],
"section": chunk.get("section", ""),
"chunk_index": chunk.get("chunk_index", 0),
"total_chunks": chunk.get("total_chunks", 1),
"pages": chunk.get("pages", []),
"token_count": chunk.get("token_count", 0),
"text": chunk["text"],
"embedding": emb,
"source_file": chunk.get("metadata", {}).get("source_file", ""),
"created_at": datetime.utcnow(),
}
ops.append(
UpdateOne(
{"document_id": document_id, "chunk_id": chunk["chunk_id"]},
{"$set": doc},
upsert=True,
)
)
if ops:
result = col.bulk_write(ops, ordered=False)
logger.info(
f"Upserted {result.upserted_count + result.modified_count} chunks."
)
print("ingested chunks in ragService")
return len(ops)
def delete_document_chunks(document_id: str) -> int:
col = get_collection()
result = col.delete_many({"document_id": document_id})
return result.deleted_count
# Retrieval
def retrieve_chunks(
query: str, document_id: str, top_k: int = TOP_K
) -> List[Dict]:
col = get_collection()
q_emb = embed_query(query)
pipeline = [
{
"$vectorSearch": {
"index": VECTOR_INDEX,
"path": "embedding",
"queryVector": q_emb,
"numCandidates": top_k * 4,
"limit": top_k,
"filter": {"document_id": {"$eq": document_id}},
}
},
{
"$project": {
"_id": 0,
"text": 1,
"section": 1,
"pages": 1,
"chunk_index": 1,
"score": {"$meta": "vectorSearchScore"},
}
},
]
try:
results = list(col.aggregate(pipeline))
#removing low confidence chunk
results = [r for r in results if r.get("score", 0) >= MIN_SCORE]
results = [
r for r in results if len(r.get("text", "")) <= MAX_CHUNK_CHARS
]
if not results:
logger.warning(
"All retrieved chunks failed quality gate "
"(low score or oversized). Falling back to text search."
)
return _fallback_text_search(query, document_id, top_k)
logger.info(
f"Retrieved {len(results)} chunk(s) "
f"(score ≥ {MIN_SCORE}, size ≤ {MAX_CHUNK_CHARS} chars)"
)
return results
except Exception as e:
logger.error(f"Vector search failed: {e}")
return _fallback_text_search(query, document_id, top_k)
def _fallback_text_search(
query: str, document_id: str, top_k: int
) -> List[Dict]:
"""Keyword fallback when the vector index isn't ready yet."""
col = get_collection()
words = query.split()[:5]
regex_part = "|".join(words)
docs = list(
col.find(
{
"document_id": document_id,
"text": {"$regex": regex_part, "$options": "i"},
},
{"_id": 0, "text": 1, "section": 1, "pages": 1, "chunk_index": 1},
).limit(top_k)
)
logger.info(f"Fallback text search returned {len(docs)} result(s).")
return docs
# sContext Builder
def build_context(
chunks: List[Dict], max_chars: int = MAX_CONTEXT_CHARS
) -> str:
"""
Build a compact context string from retrieved chunks.
Each chunk is individually truncated, then the total is capped at max_chars.
"""
if not chunks:
return "No relevant context found."
parts = []
total_chars = 0
for i, c in enumerate(chunks, 1):
section = c.get("section", "Unknown Section")
pages = c.get("pages", [])
page_str = f"p.{pages[0]}" if pages else ""
text = c["text"].strip()
# Truncate large chunks
if len(text) > CHUNK_TRUNCATE:
text = text[:CHUNK_TRUNCATE] + "…[truncated]"
entry = f"[{i} | {section} {page_str}]\n{text}"
if total_chars + len(entry) > max_chars:
logger.info(f"Context budget reached at chunk {i}, stopping.")
break
parts.append(entry)
total_chars += len(entry)
logger.info(f"Built context: {total_chars} chars from {len(parts)} chunk(s)")
return "\n\n---\n\n".join(parts)
GEMINI_MODEL = os.getenv("GEMINI_MODEL", "gemini-2.5-flash-lite")
def _get_gemini() -> genai.GenerativeModel:
if not GEMINI_API_KEY:
raise ValueError("GEMINI_API_KEY not set in environment.")
genai.configure(api_key=GEMINI_API_KEY)
return genai.GenerativeModel(GEMINI_MODEL)
SYSTEM_PROMPT = (
"Answer using ONLY the document context below. "
"If the answer isn't there, say: \"I couldn't find that in this document.\" "
"Cite section name or page number when available. "
"Be concise; use bullet points for multi-part answers."
)
def generate_answer(
query: str,
context: str,
chat_history: Optional[List[Dict]] = None,
) -> str:
"""
Call Gemini with retrieved context + optional chat history.
Retries up to MAX_RETRIES times with exponential backoff on 429 errors.
"""
model = _get_gemini()
history_text = ""
if chat_history:
recent = chat_history[-(MAX_HISTORY_TURNS * 2):]
for turn in recent:
role = "User" if turn["role"] == "user" else "Assistant"
content = turn["content"][:HISTORY_MSG_CAP]
history_text += f"{role}: {content}\n"
sections = [f"CONTEXT:\n{context}"]
if history_text:
sections.append(f"RECENT CHAT:\n{history_text.strip()}")
sections.append(f"QUESTION: {query}")
prompt = SYSTEM_PROMPT + "\n\n" + "\n\n".join(sections) + "\n\nANSWER:"
estimated_tokens = len(prompt) // 4
logger.info(f"Prompt estimate: ~{estimated_tokens} tokens | model: {GEMINI_MODEL}")
generation_config = genai.GenerationConfig(
max_output_tokens=MAX_OUTPUT_TOKENS,
temperature=TEMPERATURE,
)
# retry
last_error = None
for attempt in range(MAX_RETRIES):
try:
response = model.generate_content(prompt, generation_config=generation_config)
return response.text.strip()
except Exception as e:
last_error = e
err_str = str(e)
is_rate_limit = "429" in err_str or "quota" in err_str.lower()
is_transient = any(
code in err_str for code in ("500", "502", "503", "504")
)
if (is_rate_limit or is_transient) and attempt < MAX_RETRIES - 1:
wait = RETRY_BASE_WAIT * (2 ** attempt) # 1s → 2s → 4s
logger.warning(
f"Gemini {'rate limit' if is_rate_limit else 'transient error'} "
f"(attempt {attempt + 1}/{MAX_RETRIES}). "
f"Retrying in {wait:.0f}s… [{err_str[:80]}]"
)
time.sleep(wait)
else:
# Non-retryable error or out of attempts
logger.error(f"Gemini error after {attempt + 1} attempt(s): {e}")
raise
# Should never reach here, but satisfy type checker
raise RuntimeError(f"Gemini call failed after {MAX_RETRIES} attempts: {last_error}")
# Main RAG Pipeline
def rag_query(
query: str,
document_id: str,
chat_history: Optional[List[Dict]] = None,
) -> Dict:
"""
Full RAG pipeline: retrieve → quality gate → build context → generate answer.
Args:
query : user question
document_id : MongoDB document identifier (used to filter chunks)
chat_history: list of {"role": "user"|"assistant", "content": str}
Returns:
{
"answer": str,
"sources": [{"section": str, "pages": list, "score": float}],
"chunks_used": int,
}
"""
chunks = retrieve_chunks(query, document_id)
context = build_context(chunks)
answer = generate_answer(query, context, chat_history)
sources = [
{
"section": c.get("section", ""),
"pages": c.get("pages", []),
"score": round(c.get("score", 0), 3),
}
for c in chunks
]
return {
"answer": answer,
"sources": sources,
"chunks_used": len(chunks),
}