Files
rag-llm/main.py
2025-12-30 03:24:40 +03:00

411 lines
14 KiB
Python

#!/usr/bin/env python3
import os
import sys
import json
import hashlib
import asyncio
import re
from pathlib import Path
from collections import deque
from typing import List, Dict, Tuple
import torch
from dotenv import load_dotenv
from rich.console import Console
from rich.panel import Panel
from rich.markdown import Markdown
from prompt_toolkit import PromptSession
from prompt_toolkit.styles import Style
from prompt_toolkit.patch_stdout import patch_stdout
from langchain_community.document_loaders import UnstructuredMarkdownLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_ollama import OllamaEmbeddings, ChatOllama
from langchain_chroma import Chroma
from langchain_core.documents import Document
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from watchdog.observers import Observer
from watchdog.events import FileSystemEventHandler
# =========================
# CONFIG
# =========================
console = Console()
session = PromptSession()
load_dotenv()
style = Style.from_dict({"prompt": "bold #6a0dad"})
# --- PROMPTS ---
SYSTEM_PROMPT_SEARCH = os.getenv("SYSTEM_PROMPT", "You are a precise technical assistant. Cite sources using [filename]. Be concise.")
SYSTEM_PROMPT_ANALYSIS = (
"You are an expert tutor and progress evaluator. "
"You have access to the student's entire knowledge base below. "
"Analyze the coverage, depth, and connections in the notes. "
"Identify what the user has learned well, what is missing, and suggest the next logical steps. "
"Do not just summarize; evaluate the progress."
)
USER_PROMPT_TEMPLATE = os.getenv("USER_PROMPT_TEMPLATE",
"Previous Conversation:\n{history}\n\nContext from Docs:\n{context}\n\nCurrent Question: {question}")
MD_DIRECTORY = os.getenv("MD_FOLDER", "./notes")
EMBEDDING_MODEL = os.getenv("EMBEDDING_MODEL", "nomic-embed-text")
LLM_MODEL = os.getenv("LLM_MODEL", "llama3")
CHROMA_PATH = "./.cache/chroma_db"
HASH_CACHE = "./.cache/file_hashes.json"
MAX_EMBED_CHARS = 380
CHUNK_SIZE = 1200
CHUNK_OVERLAP = 200
TOP_K = 6
COLLECTION_NAME = "md_rag"
# Limit context size for Analysis mode (approx 24k chars ~ 6k tokens) to prevent OOM
MAX_ANALYSIS_CONTEXT_CHARS = 24000
BATCH_SIZE = 10
MAX_PARALLEL_FILES = 3
# =========================
# GPU SETUP
# =========================
def setup_gpu():
if torch.cuda.is_available():
torch.cuda.set_per_process_memory_fraction(0.95)
device_id = torch.cuda.current_device()
device_name = torch.cuda.get_device_name(device_id)
total_vram = torch.cuda.get_device_properties(device_id).total_memory / (1024**3)
console.print(f"[green]✓ GPU: {device_name} ({total_vram:.1f}GB)[/green]\n")
else:
console.print("[yellow]⚠ CPU mode[/yellow]\n")
console.print("\n")
setup_gpu()
# =========================
# UTILS & CACHE
# =========================
def get_file_hash(file_path: str) -> str:
return hashlib.md5(Path(file_path).read_bytes()).hexdigest()
def load_hash_cache() -> dict:
Path(HASH_CACHE).parent.mkdir(parents=True, exist_ok=True)
if Path(HASH_CACHE).exists():
return json.loads(Path(HASH_CACHE).read_text())
return {}
def save_hash_cache(cache: dict):
Path(HASH_CACHE).write_text(json.dumps(cache, indent=2))
# =========================
# ROUTING LOGIC
# =========================
def classify_intent(query: str) -> str:
"""
Determines if the user wants a specific search (RAG) or a global assessment.
"""
analysis_keywords = [
r"assess my progress", r"eval(uate)? my (learning|knowledge)",
r"what have i learned", r"summary of (my )?notes",
r"my progress", r"learning path", r"knowledge gap",
r"оцени (мой )?прогресс", r"что я выучил", r"итоги", r"анализ знаний"
]
query_lower = query.lower()
for pattern in analysis_keywords:
if re.search(pattern, query_lower):
return "ANALYSIS"
return "SEARCH"
# =========================
# DOCUMENT PROCESSING
# =========================
def validate_chunk_size(text: str, max_chars: int = MAX_EMBED_CHARS) -> List[str]:
if len(text) <= max_chars:
return [text]
sentences = text.replace('. ', '.|').replace('! ', '!|').replace('? ', '?|').split('|')
chunks = []
current = ""
for sentence in sentences:
if len(current) + len(sentence) <= max_chars:
current += sentence
else:
if current: chunks.append(current.strip())
# Handle extremely long sentences by word splitting
if len(sentence) > max_chars:
words = sentence.split()
temp = ""
for word in words:
if len(temp) + len(word) + 1 <= max_chars:
temp += word + " "
else:
if temp: chunks.append(temp.strip())
temp = word + " "
if temp: chunks.append(temp.strip())
current = ""
else:
current = sentence
if current: chunks.append(current.strip())
return [c for c in chunks if c]
class ChunkProcessor:
def __init__(self, vectorstore):
self.vectorstore = vectorstore
self.semaphore = asyncio.Semaphore(MAX_PARALLEL_FILES)
async def process_file(self, file_path: str) -> List[Dict]:
try:
docs = await asyncio.to_thread(UnstructuredMarkdownLoader(file_path).load)
except Exception as e:
console.print(f"[red]✗ {Path(file_path).name}: {e}[/red]")
return []
splitter = RecursiveCharacterTextSplitter(
chunk_size=CHUNK_SIZE,
chunk_overlap=CHUNK_OVERLAP,
separators=["\n\n", "\n", ". ", " "]
)
chunks = []
for doc_idx, doc in enumerate(docs):
for chunk_idx, text in enumerate(splitter.split_text(doc.page_content)):
safe_texts = validate_chunk_size(text)
for sub_idx, safe_text in enumerate(safe_texts):
chunks.append({
"id": f"{file_path}::{doc_idx}::{chunk_idx}::{sub_idx}",
"text": safe_text,
"metadata": {"source": file_path, **doc.metadata}
})
return chunks
async def embed_batch(self, batch: List[Dict]) -> bool:
if not batch: return True
try:
docs = [Document(page_content=c["text"], metadata=c["metadata"]) for c in batch]
ids = [c["id"] for c in batch]
await asyncio.to_thread(self.vectorstore.add_documents, docs, ids=ids)
return True
except Exception as e:
console.print(f"[red]✗ Embed error: {e}[/red]")
return False
async def index_file(self, file_path: str, cache: dict) -> bool:
async with self.semaphore:
current_hash = get_file_hash(file_path)
if cache.get(file_path) == current_hash:
return False
chunks = await self.process_file(file_path)
if not chunks: return False
try:
self.vectorstore._collection.delete(where={"source": file_path})
except:
pass # Collection might be empty
for i in range(0, len(chunks), BATCH_SIZE):
batch = chunks[i:i + BATCH_SIZE]
await self.embed_batch(batch)
cache[file_path] = current_hash
console.print(f"[green]✓ {Path(file_path).name} ({len(chunks)} chunks)[/green]")
return True
# =========================
# FILE WATCHER
# =========================
class DocumentWatcher(FileSystemEventHandler):
def __init__(self, processor, cache):
self.processor = processor
self.cache = cache
self.queue = deque()
self.processing = False
def on_modified(self, event):
if not event.is_directory and event.src_path.endswith(".md"):
self.queue.append(event.src_path)
async def process_queue(self):
while True:
if self.queue and not self.processing:
self.processing = True
file_path = self.queue.popleft()
if Path(file_path).exists():
await self.processor.index_file(file_path, self.cache)
save_hash_cache(self.cache)
self.processing = False
await asyncio.sleep(1)
def start_watcher(processor, cache):
handler = DocumentWatcher(processor, cache)
observer = Observer()
observer.schedule(handler, MD_DIRECTORY, recursive=True)
observer.start()
asyncio.create_task(handler.process_queue())
return observer
# =========================
# RAG CHAIN FACTORY
# =========================
class ConversationMemory:
def __init__(self, max_messages: int = 8):
self.messages = []
self.max_messages = max_messages
def add(self, role: str, content: str):
self.messages.append({"role": role, "content": content})
if len(self.messages) > self.max_messages:
self.messages.pop(0)
def get_history(self) -> str:
if not self.messages: return "No previous conversation."
return "\n".join([f"{m['role'].upper()}: {m['content']}" for m in self.messages])
def get_chain(system_prompt):
llm = ChatOllama(model=LLM_MODEL, temperature=0.2)
prompt = ChatPromptTemplate.from_messages([
("system", system_prompt),
("human", USER_PROMPT_TEMPLATE)
])
return prompt | llm | StrOutputParser()
# =========================
# MAIN
# =========================
async def main():
Path(MD_DIRECTORY).mkdir(parents=True, exist_ok=True)
Path(CHROMA_PATH).parent.mkdir(parents=True, exist_ok=True)
console.print(Panel.fit(
f"[bold cyan]⚡ Dual-Mode RAG System[/bold cyan]\n"
f"📂 Docs: {MD_DIRECTORY}\n"
f"🧠 Embed: {EMBEDDING_MODEL}\n"
f"🤖 LLM: {LLM_MODEL}",
border_style="cyan"
))
embeddings = OllamaEmbeddings(model=EMBEDDING_MODEL)
vectorstore = Chroma(
collection_name=COLLECTION_NAME,
persist_directory=CHROMA_PATH,
embedding_function=embeddings
)
processor = ChunkProcessor(vectorstore)
cache = load_hash_cache()
console.print("\n[yellow]Checking documents...[/yellow]")
files = [
os.path.join(root, file)
for root, _, files in os.walk(MD_DIRECTORY)
for file in files if file.endswith(".md")
]
# Initial Indexing
semaphore = asyncio.Semaphore(MAX_PARALLEL_FILES)
async def sem_task(fp):
async with semaphore:
return await processor.index_file(fp, cache)
tasks = [sem_task(fp) for fp in files]
for fut in asyncio.as_completed(tasks):
await fut
save_hash_cache(cache)
observer = start_watcher(processor, cache)
memory = ConversationMemory()
console.print("[bold green]💬 Ready! Type 'exit' to quit.[/bold green]\n")
try:
with patch_stdout():
while True:
query = await session.prompt_async("> ", style=style)
query = query.strip()
if query.lower() in {"exit", "quit", "q"}:
print("Goodbye!")
break
if not query: continue
mode = classify_intent(query)
history_str = memory.get_history()
if mode == "SEARCH":
console.print("[bold blue]🔍 SEARCH MODE (Top-K)[/bold blue]")
# Standard RAG
retriever = vectorstore.as_retriever(search_kwargs={"k": TOP_K})
docs = await asyncio.to_thread(retriever.invoke, query)
context_str = "\n\n".join(f"[{Path(d.metadata['source']).name}]\n{d.page_content}" for d in docs)
chain = get_chain(SYSTEM_PROMPT_SEARCH)
else: # ANALYSIS MODE
console.print("[bold magenta]📊 ANALYSIS MODE (Full Context)[/bold magenta]")
# Fetch ALL documents (limited by size)
# Chroma .get() returns dict with keys: ids, embeddings, documents, metadatas
db_data = await asyncio.to_thread(vectorstore.get)
all_texts = db_data['documents']
all_metas = db_data['metadatas']
if not all_texts:
console.print("[red]No documents found to analyze![/red]")
continue
# Concatenate content for analysis
full_context = ""
char_count = 0
# Sort arbitrarily or by source to group files
paired = sorted(zip(all_texts, all_metas), key=lambda x: x[1]['source'])
for text, meta in paired:
entry = f"\n---\nSource: {Path(meta['source']).name}\n{text}\n"
if char_count + len(entry) > MAX_ANALYSIS_CONTEXT_CHARS:
full_context += "\n[...Truncated due to context limit...]"
console.print("[yellow]⚠ Context limit reached, truncating analysis data.[/yellow]")
break
full_context += entry
char_count += len(entry)
context_str = full_context
chain = get_chain(SYSTEM_PROMPT_ANALYSIS)
response = ""
console.print(f"[dim]Context size: {len(context_str)} chars[/dim]")
async for chunk in chain.astream({
"context": context_str,
"question": query,
"history": history_str
}):
print(chunk, end="")
response += chunk
console.print("\n")
memory.add("user", query)
memory.add("assistant", response)
finally:
observer.stop()
observer.join()
if __name__ == "__main__":
import nest_asyncio
nest_asyncio.apply()
try:
import asyncio
loop = asyncio.get_event_loop()
loop.run_until_complete(main())
except KeyboardInterrupt:
console.print("\n[yellow]Goodbye![/yellow]")
sys.exit(0)