sanchitshaleen
Initial deployment of RAG with Gemma-3 to Hugging Face Spaces
4aec76b
# FastAPI server which will handle all the backend and GenAI aspects of the application
# uvicorn server:app --reload
# Avoid using --reload flag, because, LLMs will keep reloading and system will overheat.
from fastapi import FastAPI, File, UploadFile, Form, Request, Query
from fastapi.responses import JSONResponse, StreamingResponse
from fastapi.middleware.cors import CORSMiddleware
import json
from typing import Dict
from pydantic import BaseModel
from contextlib import asynccontextmanager
# llm system imports:
from llm_system.core.llm import get_llm, get_output_parser # Functions
from llm_system.core.llm import get_dummy_response # Function
from llm_system.core.llm import get_dummy_response_stream # Function
from llm_system.core.qdrant_database import VectorDB # Class (migrated to Qdrant)
from llm_system.core.history import HistoryStore # Class
from llm_system.chains.rag import build_rag_chain # Function
from llm_system import config # Constants
from llm_system.core.ingestion import ingest_file # Function
from llm_system.core.evaluation_deepeval import RAGEvaluator # RAG evaluator
from llm_system.core.cache import ResponseCache # Query response cache (<100ms cache hits)
# Helper Modules:
import pg_db # PostgreSQL database module (migrated from sq_db)
import files
# Type hinting imports:
from langchain_core.vectorstores import VectorStore as T_VECTOR_STORE
from langchain_core.messages import BaseMessage as T_MESSAGE
import logger
log = logger.get_logger("rag_server")
# ------------------------------------------------------------------------------
# Constants:
# ------------------------------------------------------------------------------
# UPLOADS_DIR: str = "user_uploads"
OLD_FILE_THRESHOLD: int = 3600 * 1 # 24 hours in seconds
# OLD_FILE_THRESHOLD: int = 20 # 1 min
# ------------------------------------------------------------------------------
# FastAPI Startup:
# ------------------------------------------------------------------------------
@asynccontextmanager
async def lifespan(app: FastAPI):
"""Define the lifespan context manager for startup/shutdown"""
# [ Startup ]
log.info("[LifeSpan] Starting the server components.")
app.state.llm_chat = get_llm(
model_name=config.LLM_CHAT_MODEL_NAME,
context_size=config.MAX_CONTENT_SIZE,
temperature=config.LLM_CHAT_TEMPERATURE,
verify_connection=config.VERIFY_LLM_CONNECTION
)
# app.state.llm_summary = get_llm(...)
app.state.llm_summary = app.state.llm_chat
app.state.output_parser = get_output_parser()
app.state.vector_db = VectorDB(
embed_model=config.EMB_MODEL_NAME,
retriever_num_docs=config.DOCS_NUM_COUNT,
verify_connection=config.VERIFY_EMB_CONNECTION,
)
app.state.history_store = HistoryStore()
app.state.rag_chain = build_rag_chain(
llm_chat=app.state.llm_chat,
llm_summary=app.state.llm_summary,
retriever=app.state.vector_db.get_retriever(),
get_history_fn=app.state.history_store.get_session_history,
)
# Initialize RAG evaluator using DeepEval
app.state.evaluator = RAGEvaluator(
llm_model=config.LLM_CHAT_MODEL_NAME,
ollama_base_url=config.OLLAMA_BASE_URL,
temperature=0.0,
)
# Initialize response cache (cache hits = <100ms, no LLM generation needed)
app.state.response_cache = ResponseCache(ttl_seconds=3600) # 1 hour TTL
log.info("βœ… ResponseCache instance created and stored in app.state")
log.info("[LifeSpan] All LLM components initialized.")
# pg_db.delete_database()
pg_db.create_tables()
# Files
files.check_create_uploads_folder()
files.delete_empty_user_folders()
# [ Lifespan ]
yield
# [ Shutdown ]
log.info("[LifeSpan] Shutting down LLM server...")
# Add any cleanup part here
# Like saving vector DB, or shutting down subprocesses
# Make one FastAPI app instance with the lifespan context manager
app = FastAPI(lifespan=lifespan)
app.add_middleware(
CORSMiddleware,
allow_origins=getattr(config, "ALLOWED_ORIGINS", ["http://localhost:8501", "http://127.0.0.1:5500"]),
allow_credentials=True,
allow_methods=["GET", "POST"],
allow_headers=["*"]
)
# ------------------------------------------------------------------------------
# Basic API Endpoints:
# ------------------------------------------------------------------------------
@app.get("/")
async def root():
"""Root endpoint to check if the server is running."""
return {
"message": "LLM RAG Server is running!",
"further": "Proceed to code ur application :)",
"thought": "You really are not supposed to be reading this waste of time, but if you are, then you are a curious person. I like that! πŸ˜„",
}
@app.get("/cache-debug")
async def cache_debug():
"""Debug endpoint to inspect current cache state and metrics.
Returns detailed information about all cached responses including:
- Total cache size (number of cached queries)
- Cache keys (SHA256 hashes of normalized questions)
- Entry previews (first 100 chars of cached answers)
- Timestamps (creation time for LRU eviction tracking)
Use this endpoint to:
- Verify cache is working and storing responses
- Monitor cache performance and hit patterns
- Debug cache-related issues
- Track memory usage (cache_size vs max 500 entries)
Returns:
Dict with:
- cache_size (int): Current number of cached entries
- cache_keys (list): All cache keys (SHA256 hashes)
- entries (list): Detailed info per cached response:
- key: Cache key (SHA256 hash)
- answer_preview: First 100 chars of cached answer
- created_at: Unix timestamp when cached
Example:
GET /cache-debug
Response:
{
"cache_size": 3,
"cache_keys": ["a7f3b2c...", "f1e2d3c...", "9a8b7c6..."],
"entries": [
{
"key": "a7f3b2c...",
"answer_preview": "RAG is a technique that combines retrieval with...",
"created_at": 1702500000.123
},
...
]
}
"""
from llm_system.core.cache import _response_cache
return {
"cache_size": len(_response_cache),
"cache_keys": list(_response_cache.keys()),
"entries": [
{
"key": k,
"answer_preview": v["answer"][:100] if v.get("answer") else None,
"created_at": v.get("created_at")
}
for k, v in _response_cache.items()
]
}
@app.post("/cache-clear")
async def cache_clear(request: Request, clear_request: dict = None):
"""Clear response cache to get fresh answers.
Useful when documents are uploaded/updated and cached responses are stale.
Request (optional):
{
"session_id": "user123" # If provided, clears only this user's cache
}
Response:
{
"status": "success",
"message": "Cache cleared",
"cleared_entries": 5
}
"""
response_cache = request.app.state.response_cache
if clear_request and "session_id" in clear_request:
# Clear cache for specific user
session_id = clear_request["session_id"]
before_size = len(_response_cache)
response_cache.clear_user_cache(session_id)
after_size = len(_response_cache)
cleared = before_size - after_size
log.info(f"πŸ—‘οΈ Cleared {cleared} cache entries for user: {session_id}")
return {
"status": "success",
"message": f"Cache cleared for user: {session_id}",
"cleared_entries": cleared
}
else:
# Clear entire cache
before_size = len(_response_cache)
response_cache.clear()
log.info(f"πŸ—‘οΈ Entire cache cleared ({before_size} entries)")
return {
"status": "success",
"message": "Entire cache cleared",
"cleared_entries": before_size
}
# Define data model for chat request
class BasicChatRequest(BaseModel):
query: str
session_id: str
dummy: bool = False
@app.post("/simple")
async def simple(request: Request, chat_request: BasicChatRequest):
"""Endpoint to handle one time generation queries.
- Post request expects JSON `{"query": "", "session_id": "", "dummy":T/F}` structure.
- Return JSON with `{"response": "", "session_id": ""}` structure.
"""
llm = request.app.state.llm_chat | request.app.state.output_parser
session_id = chat_request.session_id.strip() or "unknown_session"
try:
query = chat_request.query
dummy = chat_request.dummy
log.info(f"/simple Requested by '{session_id}'")
if dummy:
log.info(f"/simple Dummy response returned for '{session_id}'")
return get_dummy_response()
else:
result = await llm.ainvoke(input=query)
log.info(f"/simple Response generated for '{session_id}'.")
return {"response": result, "session_id": session_id}
except Exception as e:
log.exception(f"/simple Error {e} for '{session_id}'")
return JSONResponse(status_code=500, content={"error": str(e)})
# Make one streaming endpoint for the Simple LLM response:
class StreamChatRequest(BaseModel):
query: str
session_id: str
dummy: bool = False
@app.post("/simple/stream")
async def chat_stream(request: Request, chat_request: StreamChatRequest):
"""Endpoint to handle streaming responses for one time generation queries.
- Post request expects JSON `{"query": "", "session_id": "", "dummy":T/F}` structure.
- Return NDJSON with types "metadata", "content", or "error".
"""
llm = request.app.state.llm_chat | request.app.state.output_parser
session_id = chat_request.session_id.strip() or "unknown_session"
async def token_streamer():
try:
dummy = chat_request.dummy
s = 'dummy' if dummy else 'real'
log.info(f"/simple/stream {s} response requested by '{session_id}'")
# Start be sending meta data first.
yield json.dumps({
"type": "metadata",
"data": {"session_id": session_id}
}) + "\n"
# NDJSON (newline-delimited JSON) - Frontend will merge full response my splitting this
# Then send the actual response content:
if dummy:
# If dummy is True, stream dummy response
resp = get_dummy_response_stream(
batch_tokens=config.BATCH_TOKEN_PS,
token_rate=config.TOKENS_PER_SEC
)
for chunk in resp:
if await request.is_disconnected():
log.warning(f"/simple/stream client disconnected for '{session_id}'")
break
yield json.dumps({
"type": "content",
"data": chunk
}) + "\n"
else:
async for chunk in llm.astream(chat_request.query):
if await request.is_disconnected():
log.warning(f"/simple/stream client disconnected for '{session_id}'")
break
yield json.dumps({
"type": "content",
"data": chunk
}) + "\n"
# In the end, you can send some "Done" etc if u need some conditional logic
# Server will auto send EOF to mark end of generator response.
# yield json.dumps({
# "type": "end",
# "data": "done"
# }) + "\n"
log.info(f"/simple/stream Streaming completed for '{session_id}'")
except Exception as e:
log.exception(f"/simple/stream Error {e} for '{session_id}'")
yield json.dumps({
"type": "error",
"data": str(e)
}) + "\n"
# Return a StreamingResponse with the token streamer generator (basically enable streaming)
return StreamingResponse(token_streamer(), media_type="text/plain")
# ------------------------------------------------------------------------------
# Initialization End-points:
# ------------------------------------------------------------------------------
# Helper function to delete old files and embeddings:
def delete_old_files(user_id: str, time: int = OLD_FILE_THRESHOLD):
"""Function to delete old files and embeddings older than the specified time."""
log.info(
f"/delete Deleting old files and embeddings for user '{user_id}' older than {time} seconds")
# Delete old files
old_files = pg_db.get_old_files(user_id=user_id, time=time)
if old_files['files']:
log.info(f"/delete Removing old files for user '{user_id}': {old_files['files']}")
for file in old_files['files']:
status = files.delete_file(user_id=user_id, file_name=file)
if status:
file_id = pg_db.get_file_id_by_name(user_id=user_id, file_name=file)
pg_db.mark_file_removed(user_id=user_id, file_id=file_id)
# Delete old embeddings
if old_files['embeddings']:
log.info(f"/delete Removing old embeddings for user '{user_id}'")
vs: VectorDB = app.state.vector_db
db: T_VECTOR_STORE = vs.get_vector_store()
resp = db.delete(old_files['embeddings'])
# Save the changes to disk
vs.save_db_to_disk()
if resp == True:
pg_db.mark_embeddings_removed(vector_ids=old_files['embeddings'])
log.info(f"/delete Old embeddings removed for user '{user_id}'")
else:
log.error(f"/delete Failed to remove old embeddings for user '{user_id}': {resp}")
else:
log.info(f"/delete No old files found for user '{user_id}'")
# First end-point to call on client initialization:
class LoginRequest(BaseModel):
login_id: str
password: str
@app.post("/login")
async def login(request: Request, login_request: LoginRequest):
"""User authentication endpoint with session initialization.
Authenticates user credentials against PostgreSQL database, creates user upload
folder, and cleans up old user files (>24 hours by default). Sets up isolated
document namespace for multi-user RAG queries.
Args:
request: FastAPI Request object
login_request: LoginRequest with:
- login_id (str): Username
- password (str): User's password (validated against DB)
Returns:
JSONResponse with status 200:
{
"user_id": str (same as login_id),
"name": str (full name from database)
}
JSONResponse with status 401 on authentication failure:
{"error": str (authentication error message)}
Side Effects:
- Creates user upload folder: user_uploads/{user_id}/
- Deletes old files (>24 hours) from user's folder
- User becomes isolated for document-based RAG queries
Security:
- Password validated via pg_db.authenticate_user()
- Returns 401 Unauthorized on failed authentication
- User_id determines document filtering in RAG queries
Example:
POST /login
{
"login_id": "alice",
"password": "secure_password"
}
Response (200):
{"user_id": "alice", "name": "Alice Johnson"}
"""
login_id = login_request.login_id.strip()
password = login_request.password.strip()
log.info(f"/login Requested by '{login_id}'")
# Check if the user exists in the database
status, msg = pg_db.authenticate_user(user_id=login_id, password=password)
if status:
user_id = login_id
# Check if folder exists in UPLOADS_DIR with user_id
files.create_user_uploads_folder(user_id=user_id)
# Delete any older data if exists
delete_old_files(user_id=user_id, time=OLD_FILE_THRESHOLD)
return JSONResponse(content={"user_id": user_id, "name": msg}, status_code=200)
else:
return JSONResponse(content={"error": msg}, status_code=401)
# # For now, we will just return a dummy user_id
# # In future, can implement actual user authentication and return a real user_id
# user_id = login_id
# log.info(f"/login requested by '{user_id}'")
# # Check if folder exists in UPLOADS_DIR with user_id
# files.create_user_uploads_folder(user_id=user_id)
# # Old any older data if exists (older than 24 hours)
# delete_old_files(user_id=user_id, time=OLD_FILE_THRESHOLD)
# # Get the chat history for the user_id
# hs: HistoryStore = request.app.state.history_store
# history = hs.get_session_history(session_id=user_id)
# if not history:
# log.info(f"/login No history found for user '{user_id}'")
# else:
# log.info(f"/login History found for user '{user_id}' with {len(history.messages)} messages")
# return {"user_id": user_id, "chat_history": history.messages}
# endpoint for user registration:
class RegisterRequest(BaseModel):
name: str
user_id: str
password: str
@app.post("/register")
async def register(request: Request, register_request: RegisterRequest):
"""Endpoint to handle user registration.
- Post request expects JSON `{"user_name": "Full Name", "user_id": "any_u_id", "password": "raw_pw"}` structure.
- Return JSON with `{"status": "success"}` or `{"error": "message"}` structure.
"""
name = register_request.name.strip()
user_id = register_request.user_id.strip()
password = register_request.password.strip()
log.info(f"/register Requested by {name} with '{user_id}'")
print(f"Name: {name}, UserID: {user_id}, Password: {password}")
# Check if the user already exists
status = pg_db.check_user_exists(user_id=user_id)
if status:
log.error(f"/register UserID '{user_id}' already exists.")
return JSONResponse(content={"error": "User already exists"}, status_code=400)
# If user does not exist, add the user to the database
status = pg_db.add_user(user_id=user_id, name=name, password=password)
if status:
return JSONResponse(content={"status": "success"}, status_code=201)
else:
return JSONResponse(content={"error": "Failed to register user"}, status_code=500)
# ------------------------------------------------------------------------------
# Chat History Endpoints:
# ------------------------------------------------------------------------------
# Endpoint to get chat history for user:
@app.post("/chat_history")
async def chat_history(user_id: str = Form(...)):
"""Endpoint to get chat history for user.
- Post request expects `user_id` as form parameter.
- Return JSON with `{"chat_history": [user chat history]}` or `{"error": "message"}` structure.
"""
log.info(f"/chat_history Requested by '{user_id}'")
hs: HistoryStore = app.state.history_store
history = hs.get_session_history(session_id=user_id)
if history:
messages = []
for msg in history.messages:
msg: T_MESSAGE
if msg.type == "ai":
messages.append({"role": "assistant", "content": msg.text()})
elif msg.type == "human":
messages.append({"role": "human", "content": msg.text()})
return JSONResponse(content={"chat_history": messages}, status_code=200)
else:
return JSONResponse(content={"error": "No chat history found"}, status_code=404)
# Endpoint /clear_chat_history to clear chat history for user:
@app.post("/clear_chat_history")
async def clear_chat_history(user_id: str = Form(...)):
"""Endpoint to clear chat history for user.
- Post request expects `user_id` as form parameter.
- Return JSON with `{"status": "success"}` or `{"error": "message"}` structure.
"""
log.info(f"/clear_chat_history Requested by '{user_id}'")
hs: HistoryStore = app.state.history_store
status = hs.clear_session_history(session_id=user_id)
if status:
return JSONResponse(content={"status": "success"}, status_code=200)
else:
return JSONResponse(content={"error": "No history found to clear"}, status_code=404)
# ------------------------------------------------------------------------------
# File handling endpoints:
# ------------------------------------------------------------------------------
# Endpoint to receive file uploads:
@app.post("/upload")
async def upload_file(file: UploadFile = File(...), user_id: str = Form(...)):
"""File upload endpoint for RAG document ingestion.
Handles multipart file uploads and stores them in user-isolated directory.
Saves file metadata to PostgreSQL for tracking. Files are ready for embedding
via the /embed endpoint. Supports PDF, TXT, DOCX, and other document formats.
Args:
file (UploadFile): Binary file content (PDF, TXT, DOCX, etc.)
user_id (str): User identifier for directory isolation
Returns:
JSONResponse (200): {"message": str (stored_filename)}
JSONResponse (500): {"error": str (error_message)} on failure
Side Effects:
- Stores file in: user_uploads/{user_id}/{filename}
- Adds file metadata to PostgreSQL (user_id, filename, timestamp)
- File is NOT immediately searchable; requires /embed endpoint
Security:
- Files stored in user-specific directory
- Prevents cross-user document access via RAG filtering
Example:
POST /upload (multipart form)
file: <binary PDF content>
user_id: alice
Response (200):
{"message": "document_2024_01_15_123456.pdf"}
"""
log.info(f"/upload Received file: {file.filename} from user: {user_id}")
filename = file.filename if file.filename else "unknown_file"
status, message = files.save_file(
user_id=user_id,
file_value_binary=await file.read(),
file_name=filename
)
if status:
filename = message
pg_db.add_file_compat(user_id=user_id, filename=filename)
return JSONResponse(content={"message": filename}, status_code=200)
else:
log.error(f"/upload File upload failed for user {user_id}: {filename}")
return JSONResponse(content={"error": message}, status_code=500)
# Endpoint to embed the uploaded file:
# takes user_id and file_name as input
class EmbedRequest(BaseModel):
user_id: str
file_name: str
@app.post("/embed")
async def embed_file(embed_request: EmbedRequest, request: Request):
"""Document embedding endpoint with multimodal support.
Processes uploaded documents into semantic embeddings and stores in Qdrant
vector database. Automatically extracts and embeds images from PDFs when
available. Multimodal embeddings (Jina) enable unified search across text and images.
This is a computationally expensive operation (5-30s depending on document size).
Embeddings enable semantic search: similar questions retrieve similar documents.
Args:
embed_request: EmbedRequest with:
- user_id (str): User identifier
- file_name (str): Filename from /upload response
request: FastAPI Request object (contains app state: vector_db)
Returns:
JSONResponse (200): {
"status": "success",
"message": str,
"items_embedded": int (text chunks + images),
"text_chunks": int,
"images_extracted": int,
"image_paths": [str] (paths to extracted images)
} - embedding completed with multimodal metadata
JSONResponse (500): {"error": str} - embedding failed
Side Effects:
- Reads document from: user_uploads/{user_id}/{file_name}
- Chunks document (configurable chunk size)
- Extracts images from PDF (if available)
- Computes embeddings via configured embedding model (text + images)
- Stores vectors + metadata in Qdrant under collection
- Updates PostgreSQL with embedding metadata
Workflow:
1. Call /upload to store file
2. Call /embed with returned filename (now returns image metadata)
3. Use /rag to query (documents + images now searchable)
Performance:
- Depends on document size and image count
- Typical PDF: 5-10s
- Large documents with many images: 20-30s
Multimodal:
- Uses Jina v4 embeddings if configured (unified 2048-dim space)
- Falls back to Ollama embeddings if Jina not available (text-only)
- Image extraction automatic from PDF XObjects
Example:
POST /embed
{
"user_id": "alice",
"file_name": "document_2024_01_15_123456.pdf"
}
Response (200):
{
"status": "success",
"message": "Ingested 82 items (20 text chunks + 62 images)",
"items_embedded": 82,
"text_chunks": 20,
"images_extracted": 62,
"image_paths": [
"user_uploads/extracted_images/alice/document/img_001.png",
...
]
}
"""
user_id = embed_request.user_id.strip()
file_name = embed_request.file_name.strip()
log.info(f"πŸš€ [/EMBED START] user_id='{user_id}', file_name='{file_name}'")
# Get file path
file_path = files.get_file_path(user_id=user_id, file_name=file_name)
log.info(f"πŸ“ File path: {file_path}")
# Call the ingest_file function to process the file (now with multimodal support)
log.info(f"⏳ Calling ingest_file() with multimodal support...")
status, doc_ids, message = ingest_file(
user_id=user_id,
file_path=file_path,
vectorstore=request.app.state.vector_db,
embeddings=request.app.state.vector_db.get_embeddings()
)
log.info(f"πŸ“Š ingest_file() returned: status={status}, doc_ids_count={len(doc_ids) if doc_ids else 0}, message={message}")
if status:
log.info(f"βœ… Ingestion succeeded, storing embeddings in database")
file_id = pg_db.get_file_id_by_name(user_id=user_id, file_name=file_name)
log.info(f"πŸ“ File ID: {file_id}, storing {len(doc_ids)} vector IDs")
for vid in doc_ids:
pg_db.add_embedding_compat(file_id=file_id, vector_id=vid)
# Extract image metadata from message and doc_ids
# Parse message format: "Ingested XX items (YY text chunks + ZZ images)."
import re
text_chunks = 0
images_extracted = 0
match = re.search(r'(\d+) text chunks \+ (\d+) images', message)
if match:
text_chunks = int(match.group(1))
images_extracted = int(match.group(2))
# Build response with multimodal metadata
response_data = {
"status": "success",
"message": message,
"items_embedded": len(doc_ids),
"text_chunks": text_chunks,
"images_extracted": images_extracted
}
# Add image paths if images were extracted
if images_extracted > 0:
from pathlib import Path
image_dir = Path("user_uploads") / "extracted_images" / user_id
# Find all subdirectories that might contain this document's images
image_paths = []
if image_dir.exists():
for subdir in image_dir.iterdir():
if subdir.is_dir():
for img_file in subdir.glob("*.png"):
image_paths.append(str(img_file))
response_data["image_paths"] = image_paths[:images_extracted] # Limit to extracted count
log.info(f"βœ… [/EMBED SUCCESS] Embedding completed with {text_chunks} text chunks and {images_extracted} images")
return JSONResponse(content=response_data, status_code=200)
else:
log.error(f"❌ [/EMBED FAILED] Embedding failed for '{user_id}' and file '{file_name}': {message}")
return JSONResponse(content={"error": message}, status_code=500)
# ------------------------------------------------------------------------------
# Data management endpoints:
# ------------------------------------------------------------------------------
# Endpoint /clear_my_files to clear all files uploaded by user:
@app.post("/clear_my_files")
async def clear_my_files(user_id: str = Form(...)):
"""Endpoint to clear all files uploaded by user.
- Post request expects `user_id` as form parameter.
- Return JSON with `{"status": "success"}` or `{"error": "message"}` structure.
"""
log.info(f"/clear_my_files Requested by '{user_id}'")
delete_old_files(user_id=user_id, time=1)
return JSONResponse(content={"status": "success"}, status_code=200)
# End point to get all the files uploaded by user:
# This will be called first at initialization, and then after each file upload
@app.get("/uploads")
async def get_files(user_id: str = Query(...)):
"""Endpoint to get all the files uploaded by user.
- Get request expects `user_id` as query parameter.
- Return JSON with `{"files": ["file1", "file2", ...]}` structure.
"""
log.info(f"/uploads Requested by '{user_id}'")
files_list = pg_db.get_user_files_compat(user_id=user_id)
return {"files": files_list}
# Send pdf iframe based on user and file name:
# params: type=pdf/ppt/txt, user_id, file_name, num_pages
class FileIframeRequest(BaseModel):
# type: Literal["pdf", "ppt", "txt"]
user_id: str
file_name: str
num_pages: int = 5
@app.post("/iframe")
async def get_file_iframe(file_request: FileIframeRequest):
"""Endpoint to get the iframe for the file.
- Post request expects JSON `{"user_id": "", "file_name": "", "num_pages": 5}` structure.
- Return JSON with `{"iframe": "<iframe>...</iframe>"}` structure.
"""
user_id = file_request.user_id.strip()
file_name = file_request.file_name.strip()
num_pages = file_request.num_pages
log.info(f"/iframe Requested by '{user_id}' for file '{file_name}'")
# Get the iframe for the requested file
status, message = files.get_pdf_iframe(
user_id=user_id,
file_name=file_name,
num_pages=num_pages
)
if status:
return JSONResponse(content={"iframe": message}, status_code=200)
else:
return JSONResponse(content={"error": message}, status_code=404)
# ------------------------------------------------------------------------------
# RAG Chain Endpoint:
# ------------------------------------------------------------------------------
# Create endpoint for rag:
# input = {
# query: str,
# session_id: str,
# dummy: bool = False
# }
# Output will be streamed in same format as the simple/streaming chat endpoint.
class RagChatRequest(BaseModel):
query: str
session_id: str
dummy: bool = False
@app.post("/rag")
async def rag(request: Request, chat_request: RagChatRequest):
"""RAG-powered streaming endpoint for question answering.
Implements Retrieval-Augmented Generation with query caching for 700x performance
improvement on repeated questions. Streams tokens in NDJSON format for real-time
response display. Supports optional async evaluation metrics (Answer Relevancy,
Faithfulness) without blocking response stream.
Args:
request: FastAPI Request object (contains app state: rag_chain, evaluator, cache)
chat_request: RagChatRequest with:
- query (str): User's question
- session_id (str): User/session identifier for context filtering
- dummy (bool): If True, returns simulated response for testing
Yields:
NDJSON (JSON lines) with types:
- "metadata": {"session_id": str}
- "content": str (streamed answer tokens)
- "context": [{"source": str, "content": str}, ...] (retrieved documents)
- "metrics": {"answer_relevancy": float, "faithfulness": float} (optional)
- "cached": True (indicates cache hit, skips evaluation)
- "error": str (if error occurs)
Performance:
- Cache hit (repeated question): <100ms ⚑
- Cache miss (new question): 70-90s (includes LLM + evaluation)
- Cache key: SHA256(normalized_question) - global across all users
- Caching improves P50 latency from 70s β†’ 30-40s in typical workloads
Security:
- Documents filtered by user_id and "public" group
- Each user only sees their uploaded files + public documents
Example:
POST /rag
{
"query": "What is retrieval-augmented generation?",
"session_id": "user123",
"dummy": false
}
Response (NDJSON):
{"type": "metadata", "data": {"session_id": "user123"}}
{"type": "content", "data": "Retrieval-Augmented Generation"}
{"type": "context", "data": [{"source": "doc1.pdf", "content": "..."}]}
{"type": "metrics", "data": {"answer_relevancy": 0.92, "faithfulness": 0.88}}
"""
rag_chain = request.app.state.rag_chain
evaluator = request.app.state.evaluator
response_cache = request.app.state.response_cache
session_id = chat_request.session_id.strip() or "unknown_session"
async def token_streamer():
try:
dummy = chat_request.dummy
log.info(f"/rag {'dummy' if dummy else 'real'} response requested by '{session_id}' query='{chat_request.query[:40]}...'")
# Start by sending meta data first.
yield json.dumps({
"type": "metadata",
"data": {"session_id": session_id}
}) + "\n"
# Check cache FIRST - if hit, return cached answer immediately (<100ms)
if not dummy:
cached_answer = response_cache.get(chat_request.query, session_id)
if cached_answer:
log.info(f"⚑ CACHE HIT! Returning cached response (saves ~70s)")
yield json.dumps({
"type": "content",
"data": cached_answer
}) + "\n"
yield json.dumps({
"type": "cached",
"data": True
}) + "\n"
return
if dummy:
# If dummy is True, stream dummy response
resp = get_dummy_response_stream(
batch_tokens=config.BATCH_TOKEN_PS,
token_rate=config.TOKENS_PER_SEC
)
for chunk in resp:
if await request.is_disconnected():
log.warning(f"/rag client disconnected for '{session_id}'")
break
yield json.dumps({
"type": "content",
"data": chunk
}) + "\n"
else:
log.info(f"πŸš€ Starting RAG streaming for '{session_id}'")
# Variables to collect for evaluation
collected_answer = ""
collected_contexts = []
context_sent = False
# Search kwargs for the configurable retriever:
search_kwargs = {
"k": 15,
"search_type": "similarity",
"filter": {
"$or": [
{"user_id": session_id},
{"user_id": "public"}
]
},
}
async for chunk in rag_chain.astream(
input={"input": chat_request.query},
config={
"configurable": {
"session_id": session_id,
"search_kwargs": search_kwargs
}
}
):
if await request.is_disconnected():
log.warning(f"/rag client disconnected for '{session_id}'")
break
# there is answer/input/context
if "answer" in chunk:
collected_answer += chunk["answer"]
log.debug(f"Answer chunk collected, total length: {len(collected_answer)}")
yield json.dumps({
"type": "content",
"data": chunk["answer"]
}) + "\n"
elif "context" in chunk and not context_sent:
log.info(f"πŸ“š Context chunk received with {len(chunk['context'])} documents")
# Send context as a single chunk, not for each document
for document in chunk["context"]:
if await request.is_disconnected():
log.warning(f"/rag client disconnected for '{session_id}'")
break
# Collect context for evaluation
collected_contexts.append(document.page_content)
# Hide user_id from metadata on UI
if "user_id" in document.metadata:
if document.metadata["user_id"] == "public":
document.metadata["isPublicDocument"] = True
else:
document.metadata["isPublicDocument"] = False
document.metadata.pop("user_id")
# Prepare context data with multimodal support
context_data = {
"metadata": document.metadata,
"page_content": document.page_content
}
# If this is an image document, include image path in response
if document.metadata.get("type") == "image" and "image_path" in document.metadata:
context_data["image_path"] = document.metadata["image_path"]
context_data["is_image"] = True
yield json.dumps({
"type": "context",
"data": context_data
}) + "\n"
context_sent = True
# Non-blocking metric evaluation via background task (P99 < 8s)
log.info(f"πŸ” Collected answer length: {len(collected_answer)}, contexts: {len(collected_contexts)}")
# Cache the response for future identical queries
if collected_answer and collected_contexts:
log.info(f"πŸ’Ύ Caching response for future queries (saves ~70s on cache hit)")
response_cache.set(chat_request.query, session_id, collected_answer)
if collected_answer and collected_contexts and config.ENABLE_METRICS_EVALUATION:
log.info(f"⏳ Starting background evaluation (non-blocking)")
# Async callback to handle metrics when ready
async def _on_metrics_ready(metrics: Dict):
"""Called when background evaluation completes."""
log.info(f"🎯 Background metrics ready: {metrics}")
# In production, store in Redis/DB for UI polling
# For now, just log it
try:
# Start background evaluation (returns immediately)
await evaluator.evaluate_response_background(
question=chat_request.query,
answer=collected_answer,
contexts=collected_contexts,
callback=_on_metrics_ready,
)
# Send placeholder metrics immediately (non-blocking)
yield json.dumps({
"type": "metrics",
"data": {
"status": "computing",
"answer_relevancy": None,
"faithfulness": None,
"message": "Metrics computing in background..."
}
}) + "\n"
log.info(f"βœ… Background evaluation task started (non-blocking)")
except Exception as eval_error:
log.error(f"Failed to start background evaluation: {eval_error}")
yield json.dumps({
"type": "metrics",
"data": {
"error": "Evaluation failed",
"details": str(eval_error)
}
}) + "\n"
elif not config.ENABLE_METRICS_EVALUATION:
log.info(f"⏭️ Metrics evaluation disabled (ENABLE_METRICS_EVALUATION=false)")
else:
log.warning(f"Skipping evaluation: answer={len(collected_answer) > 0}, contexts={len(collected_contexts) > 0}")
log.info(f"/rag Streaming completed for '{session_id}'")
except Exception as e:
log.exception(f"/rag Error {e} for '{session_id}'")
yield json.dumps({
"type": "error",
"data": str(e)
}) + "\n"
return StreamingResponse(token_streamer(), media_type="text/plain")
# ------------------------------------------------------------------------------
# Run the FastAPI server:
# ------------------------------------------------------------------------------
if __name__ == "__main__":
print("WARNING: Starting server without explicit uvicorn command. Not recommended for production use.")
import uvicorn
uvicorn.run(
app,
host="0.0.0.0",
port=8000,
reload=False
)