Update rag_engine.py

rag_engine.py

@@ -1,341 +1,9 @@
-
-I want to work as an AI engineer but I need to implement and deploy some practical projects to show my skills on my online website portfolio and add live demo. What are the projects you suggust me to have ?
-
-01:15
-
-Reading frontend design skill for creating a good portfolio project showcase
-Great question! Let me give you a well-thought-out roadmap of projects that will genuinely impress hiring managers and demonstrate real AI engineering skills — not just API wrappers.
-
-Here's what I recommend, organized by the core skills they showcase:
-
-1. RAG-Powered Document Q&A System Build a system where users upload PDFs/docs and chat with them. This shows you understand embeddings, vector databases (Pinecone, Chroma, Qdrant), chunking strategies, retrieval pipelines, and prompt engineering. Use LangChain or LlamaIndex + a frontend. This is probably the #1 project recruiters look for.
-
-2. Fine-Tuned Model for a Niche Task Pick a specific domain (legal clause classification, medical note summarization, code review) and fine-tune an open-source model (Mistral, LLaMA, Phi) using LoRA/QLoRA on Hugging Face. Deploy it with a Gradio or FastAPI demo. This shows you can go beyond prompting and actually train/adapt models.
-
-3. Multi-Agent System Build a team of agents that collaborates on a complex task — e.g., a research agent that searches the web, a writer agent that drafts content, and an editor agent that refines it. Use LangGraph, CrewAI, or AutoGen. This signals you understand agentic architectures, tool use, and orchestration.
-
-4. Real-Time AI Voice/Chat Assistant with Memory Build a chatbot with persistent memory across sessions (using something like Mem0 or a simple vector store). Add a voice interface with Whisper for STT and ElevenLabs or TTS for speech output. This demonstrates multimodal handling and stateful AI systems.
-
-5. MLOps Pipeline with Model Monitoring Deploy a model with a full pipeline: training → evaluation → versioning (MLflow) → containerization (Docker) → deployment (FastAPI on AWS/GCP/Render) → drift monitoring. This is critical because many "AI engineers" skip this and it's a massive differentiator.
-
-6. AI-Powered Data Analysis Agent Build a natural language interface to a SQL database or CSV files. The user asks questions in plain English and gets charts + insights back. Use a Code Interpreter-style pattern. Tools: LangChain's SQL agent, Pandas, Plotly, Streamlit.
-
-7. Semantic Search Engine Build a search engine for a dataset (news articles, recipes, job listings) that uses embedding-based similarity instead of keyword search. Show a live comparison between traditional keyword search and your semantic approach. Demonstrates understanding of embeddings and information retrieval.
-
-A few strategic tips:
-
-Each project should have a clean README on GitHub, a live demo link (Hugging Face Spaces, Vercel, Render, or Streamlit Cloud are all free), and a short write-up explaining your architectural decisions and trade-offs — not just what you built, but why you made certain choices. That's what separates a strong portfolio from a list of side projects.
-
-For your portfolio website itself, lead with the live demos rather than code screenshots. Recruiters spend 30 seconds on a portfolio — make the demos one-click accessible.
-
-Would you like me to help you build out the live demo page for your portfolio, or dive deeper into how to implement any of these specific projects?
-
-now lets make this project:
-1. RAG-Powered Document Q&A System Build a system where users upload PDFs/docs and chat with them. This shows you understand embeddings, vector databases (Pinecone, Chroma, Qdrant), chunking strategies, retrieval pipelines, and prompt engineering. Use LangChain or LlamaIndex + a frontend. if it requires data , it should download it automatically and train it self. I just want to push your code to huggingface and run it live
-
-01:18
-Let me ask a few quick questions before we start building so the project is perfectly tailored for you!
-
-Q: Which vector database do you want to use?
-A: ChromaDB (local, no setup)
-
-Q: Which LLM backend do you want to power the Q&A?
-A: HuggingFace open-source model (free, no API key)
-
-Q: What kind of UI do you want?
-A: Streamlit (clean & popular)
-
-01:18
-Perfect choices! 100% free, no API keys needed, and deploys beautifully on HuggingFace Spaces. Let me build the full project now.
-
-
-Check available skills and setup workspace
-Now let me build the complete project — all files:
-
-
-Created 7 files
-
-Main Streamlit app - the RAG Q&A system
-app.py
-
-RAG Engine - the core logic: embeddings, ChromaDB, retrieval, LLM generation
-rag_engine.py
-
-Auto data downloader - downloads a sample AI report PDF for demo purposes
-data_downloader.py
-
-requirements.txt - all dependencies for the project
-requirements.txt
-
-HuggingFace Spaces README / configuration file
-README.md
-
-packages.txt for HuggingFace Spaces system dependencies
-packages.txt
-
-.gitignore to keep the repo clean
-.gitignore
-Done
-Now let's verify all files look correct:
-
-
-Verify all project files exist with correct line counts
-Now copy everything to outputs:
-
-
-Ran a command, read a file
-All 6 files are ready. Here's everything you need to deploy:
-
-📁 What was built (980 lines total)
-File	Purpose
-app.py	Streamlit UI — dark themed, chat interface, stats, sidebar
-rag_engine.py	Core RAG pipeline: chunking → embeddings → ChromaDB → MMR retrieval → Zephyr-7B
-data_downloader.py	Auto-downloads arxiv PDFs; generates a rich AI overview doc as fallback
-requirements.txt	All Python dependencies
-packages.txt	System libs for HuggingFace Spaces (poppler for PDF)
-README.md	Architecture diagram + setup instructions
-🚀 Deploy to HuggingFace Spaces (3 steps)
-Step 1 — Create the Space:
-
-Go to huggingface.co/new-space
-Name it docmind-ai, choose Streamlit, set to Public
-Click "Create Space"
-Step 2 — Push the files:
-
-bash
-git clone https://huggingface.co/spaces/YOUR_USERNAME/docmind-ai
-cd docmind-ai
-
-# Copy all 6 files into this folder, then:
-git add .
-git commit -m "🧠 DocMind AI - RAG Q&A System"
-git push
-Step 3 — (Optional) Add HF Token for higher rate limits: In your Space → Settings → Secrets → add HF_TOKEN = your HuggingFace token (free at huggingface.co/settings/tokens)
-
-HuggingFace will auto-install everything from requirements.txt and launch the app. Takes ~3 minutes on first build.
-
-App
-PY 
-Rag engine
-PY 
-Data downloader
-PY 
-Requirements
-TXT 
-Readme
-Document · MD 
-Packages
-TXT 
-1771338283280_image.png
-no option for streamlit. provide step by step guideline
-
-01:25
-I can see the issue — HuggingFace changed their UI and now shows Gradio, Docker, and Static only. Streamlit is hidden inside Docker. Here's exactly what to do:
-
-🚀 Step-by-Step: Deploy DocMind AI on HuggingFace Spaces
-Step 1 — Choose Docker (not Gradio)
-On the screen you're seeing, click Docker (the middle option with 17 templates)
-
-Step 2 — Select "Blank" template
-After clicking Docker, it will show Docker templates — select Blank
-
-Step 3 — Fill the settings
-Owner: Ryanfafa ✅ (already set)
-Space name: docmind-ai ✅ (already set)
-License: mit ✅ (already set)
-Hardware: CPU Basic (Free) ✅
-Then click "Create Space"
-
-Step 4 — We need a Dockerfile
-Since we're using Docker mode, we need one extra file. Here it is:
-
-
-Dockerfile for HuggingFace Spaces Docker deployment with Streamlit
-Also update the README header to reflect Docker SDK:
-
-
-Edited a file, read a file
-Step 5 — Clone your new Space locally
-After creating the space, open your terminal and run:
-
-bash
-git clone https://huggingface.co/spaces/Ryanfafa/docmind-ai
-cd docmind-ai
-```
-
----
-
-### Step 6 — Copy all project files into the folder
-
-You need these **7 files** total inside the `docmind-ai` folder:
-```
-docmind-ai/
-├── Dockerfile          ← NEW (just downloaded above)
-├── app.py
-├── rag_engine.py
-├── data_downloader.py
-├── requirements.txt
-├── packages.txt
-└── README.md           ← updated above
-Step 7 — Push to HuggingFace
-bash
-git add .
-git commit -m "🧠 DocMind AI - RAG Q&A with Streamlit + Docker"
-git push
-💡 If git push asks for credentials: username = your HF username, password = your HF access token (get it at huggingface.co/settings/tokens → New token → Write access)
-
-Step 8 — Watch it build
-Go to your Space URL: https://huggingface.co/spaces/Ryanfafa/docmind-ai
-
-Click the "Logs" tab to watch the Docker build. First build takes 4–6 minutes. Once it says "Running" in green — your app is live! 🎉
-
-⚠️ One known gotcha: If the build fails with a memory error, go to Space Settings → Factory reboot to try again. The free CPU Basic tier occasionally has cold-start issues.
-
-Dockerfile
- 
-Readme
-Document · MD 
-how to open terminal? I have free version
-
-01:27
-You don't need to install anything! Here are your options based on what you already have:
-
-Option 1 — Use HuggingFace's Web UI (Upload files directly, no terminal needed) ✅ EASIEST
-Go to your Space: huggingface.co/spaces/Ryanfafa/docmind-ai
-Click the "Files" tab
-Click "Add file" → "Upload files"
-Drag and drop all 7 files at once
-Click "Commit changes" — done! Build starts automatically
-Option 2 — Use VS Code in your browser (free, no install)
-Go to github.dev or vscode.dev — free browser VS Code
-Or go to any GitHub repo and press the . key to open it in browser VS Code
-Option 3 — Windows: Use the built-in terminal (no install needed)
-Press Windows key + R → type cmd → press Enter
-
-That opens Command Prompt. Then run:
-
-winget install Git.Git
-After Git installs, restart Command Prompt and you're ready.
-
-Option 4 — Mac: Terminal is already installed
-Press Cmd + Space → type Terminal → press Enter. It's built into every Mac.
-
-Which operating system are you on? (Windows / Mac / Linux) — I'll give you the exact commands for your machine, copy-paste ready.
-
-YAML Metadata Error: "short_description" length must be less than or equal to 60 characters long
-
-01:30
-Easy fix! The description is just too long. Here's the corrected README.md:
-
-
-Edited 2 files, read a file
-Download the updated README.md and re-upload it to HuggingFace, replacing the old one. The fix is just this line changed to:
-
-short_description: Chat with any PDF using RAG + ChromaDB
-That's 50 characters — safely under the 60 limit. ✅
-
-Readme
-Document · MD 
-uploaded, what next?
-
-01:31
-Now wait for the build to finish! Here's what to do:
-
-1 — Watch the build logs
-Go to your Space URL:
-
-huggingface.co/spaces/Ryanfafa/docmind-ai
-Click the "Logs" tab at the top. You'll see Docker building in real time.
-
-Normal build time: 4–6 minutes on first run.
-
-2 — What you'll see in logs
-The build goes through these stages in order:
-
-Step 1/10 — FROM python:3.10-slim → downloading base image
-Step 3/10 — pip install... → installing all packages (this is the slowest part)
-Step 7/10 — COPY . . → copying your files
-Successfully built xxxxxx → ✅ done!
-3 — When it turns green
-Once the status badge says "Running" in green, click the "App" tab — your live DocMind AI will be there!
-
-Share a screenshot of your Logs tab if you see any red error — I'll fix it immediately. What does it show right now?
-
-DocMind AI
-⚠️ Could not generate answer. Error: 410 Client Error: Gone for url: https://api-inference.huggingface.co/models/HuggingFaceH4/zephyr-7b-beta
-__https://api-inference.huggingface.co__ is no longer supported. Please use __https://router.huggingface.co__ instead.
-
-01:42
-The HuggingFace API endpoint changed! Easy fix — we just need to update the rag_engine.py to use the new router URL and a currently available free model.
-
-
-Edited a file, viewed a file
-Good, I can see the full file. Let me rewrite it completely with the fix:
-
-
-Created a file, ran a command, read a file
-Download the updated rag_engine.py and replace the one in your HuggingFace Space.
-
-What changed (2 things only)
-Old broken code:
-
-python
-# Used deprecated endpoint + old client
-from huggingface_hub import InferenceClient
-client = InferenceClient(model="HuggingFaceH4/zephyr-7b-beta")
-client.text_generation(prompt, ...)
-New fixed code:
-
-python
-# Uses new router URL + standard requests + Mistral model
-import requests
-HF_API_URL = "https://router.huggingface.co/hf-inference/models/mistralai/Mistral-7B-Instruct-v0.3/v1/chat/completions"
-requests.post(HF_API_URL, headers=headers, json=payload)
-How to update on HuggingFace
-Go to your Space → Files tab
-Click on rag_engine.py
-Click the pencil ✏️ edit icon (top right of the file)
-Select all the text and paste the new file contents
-Click "Commit changes"
-The Space will auto-rebuild in ~2 minutes and the LLM will work! ✅
-
-💡 Pro tip: Add a free HuggingFace token as a Secret (HF_TOKEN) in your Space Settings for higher rate limits and faster responses.
-
-Rag engine
-PY 
-⚠️ Could not generate answer. Error: 401 Client Error: Unauthorized for url: https://router.huggingface.co/hf-inference/models/mistralai/Mistral-7B-Instruct-v0.3/v1/chat/completions
-
-01:47
-The new HuggingFace router requires a token even for free models. Two options — let's do both: add the token AND add a guaranteed-free fallback model.
-
-Step 1 — Get your free HuggingFace token (2 minutes)
-Go to huggingface.co/settings/tokens
-Click "New token"
-Name it docmind, Role: Read
-Click "Create" → Copy the token (starts with hf_...)
-Step 2 — Add it as a Secret in your Space
-Go to your Space → Settings tab
-Scroll down to "Variables and secrets"
-Click "New secret"
-Name: HF_TOKEN → Value: paste your hf_... token
-Click Save — Space restarts automatically ✅
-Step 3 — Also update rag_engine.py with a smarter fallback
-While you're doing Step 1 & 2, download this updated file that tries multiple free models in order if one fails:
-
-
-Ran a command, read a file
-
-Write updated rag_engine.py with multi-model fallback and clear token instructions
-bash
-
-cat > /mnt/user-data/outputs/rag-document-qa/rag_engine.py << 'ENDOFFILE'
 """
 RAG Engine
-──────────
-- Embeddings : sentence-transformers/all-MiniLM-L6-v2  (HuggingFace, free)
-- Vector DB  : ChromaDB  (local, in-memory / persistent)
-- LLM        : HuggingFace Router API  (tries multiple free models)
-- Chunking   : Recursive character splitter with overlap
+Embeddings : sentence-transformers/all-MiniLM-L6-v2
+Vector DB  : ChromaDB (local, in-memory / persistent)
+LLM        : HuggingFace Router API (tries multiple free models)
+Chunking   : Recursive character splitter with overlap
 """
 
 import os
@@ -351,7 +19,7 @@ from langchain_community.embeddings import HuggingFaceEmbeddings
 from langchain_community.vectorstores import Chroma
 from langchain_community.document_loaders import PyPDFLoader, TextLoader
 
-# ─── Configuration ─────────────────────────────────────────────────────────────
+# Configuration
 EMBED_MODEL     = "sentence-transformers/all-MiniLM-L6-v2"
 CHUNK_SIZE      = 800
 CHUNK_OVERLAP   = 150
@@ -359,7 +27,7 @@ TOP_K           = 4
 COLLECTION_NAME = "docmind_collection"
 CHROMA_DIR      = "./chroma_db"
 
-# Free models to try in order (all on HF router)
+# Free models to try in order
 CANDIDATE_MODELS = [
     "mistralai/Mistral-7B-Instruct-v0.3",
     "microsoft/Phi-3.5-mini-instruct",
@@ -369,7 +37,7 @@ CANDIDATE_MODELS = [
 
 
 class RAGEngine:
-    """Full RAG pipeline: ingest → embed → store → retrieve → generate."""
+    """Full RAG pipeline: ingest, embed, store, retrieve, generate."""
 
     def __init__(self):
         self._embeddings  = None
@@ -380,7 +48,6 @@ class RAGEngine:
             separators=["\n\n", "\n", ". ", " ", ""],
         )
 
-    # ── Lazy-load embeddings ───────────────────────────────────────────────────
     @property
     def embeddings(self):
         if self._embeddings is None:
@@ -391,17 +58,15 @@ class RAGEngine:
             )
         return self._embeddings
 
-    # ── Ingest uploaded Streamlit file ────────────────────────────────────────
     def ingest_file(self, uploaded_file) -> int:
-        suffix = _get_suffix(uploaded_file.name)
+        suffix = get_suffix(uploaded_file.name)
         with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
             tmp.write(uploaded_file.read())
             tmp_path = tmp.name
         return self.ingest_path(tmp_path, uploaded_file.name)
 
-    # ── Ingest from file path ─────────────────────────────────────────────────
     def ingest_path(self, path: str, name: str = "") -> int:
-        suffix = _get_suffix(name or path)
+        suffix = get_suffix(name or path)
         loader = PyPDFLoader(path) if suffix == ".pdf" else TextLoader(path, encoding="utf-8")
         raw_docs = loader.load()
 
@@ -419,252 +84,37 @@ class RAGEngine:
         )
         return len(chunks)
 
-    # ── Query ─────────────────────────────────────────────────────────────────
     def query(self, question: str) -> Tuple[str, List[str]]:
         if self._vectorstore is None:
-            return "⚠️ Please upload a document first.", []
+            return "Please upload a document first.", []
 
         retriever = self._vectorstore.as_retriever(
             search_type="mmr",
             search_kwargs={"k": TOP_K, "fetch_k": TOP_K * 3},
         )
-        docs     = retriever.invoke(question)
-        context  = "\n\n---\n\n".join(f"[Chunk {i+1}]\n{d.page_content}" for i, d in enumerate(docs))
-        sources  = list({d.metadata.get("source", "Document") for d in docs})
-        answer   = self._generate(question, context)
-        return answer, sources
-
-    # ── LLM: try each model until one works ───────────────────────────────────
-    def _generate(self, question: str, context: str) -> str:
-        hf_token = os.environ.get("HF_TOKEN", "")
-
-        if not hf_token:
-            return (
-                "⚠️ **HF_TOKEN not set.**\n\n"
-                "To enable AI answers:\n"
-                "1. Get a free token at huggingface.co/settings/tokens\n"
-                "2. Add it as a **Secret** named `HF_TOKEN` in your Space Settings\n\n"
-                f"**Most relevant excerpt from your document:**\n\n{_extract_best(question, context)}"
-            )
-
-        headers = {
-            "Content-Type":  "application/json",
-            "Authorization": f"Bearer {hf_token}",
-        }
-        messages = [
-            {
-                "role": "system",
-                "content": (
-                    "You are DocMind, an expert document analyst. "
-                    "Answer using ONLY the provided document context. "
-                    "Be concise and cite specific details. "
-                    "If the answer is not in the context, say so clearly."
-                ),
-            },
-            {
-                "role": "user",
-                "content": f"Document context:\n{context}\n\nQuestion: {question}",
-            },
-        ]
-
-        last_error = ""
-        for model_id in CANDIDATE_MODELS:
-            try:
-                url  = f"https://router.huggingface.co/hf-inference/models/{model_id}/v1/chat/completions"
-                resp = requests.post(
-                    url,
-                    headers=headers,
-                    json={"model": model_id, "messages": messages, "max_tokens": 512, "temperature": 0.2},
-                    timeout=60,
-                )
-                if resp.status_code == 200:
-                    answer = resp.json()["choices"][0]["message"]["content"].strip()
-                    if answer:
-                        return answer
-                else:
-                    last_error = f"{resp.status_code}: {resp.text[:200]}"
-            except Exception as e:
-                last_error = str(e)
-                continue
-
-        # All models failed → extractive fallback
-        excerpt = _extract_best(question, context)
-        return (
-            f"*(All LLM models unavailable – showing most relevant excerpt)*\n\n"
-            f"{excerpt}\n\n"
-            f"<small>Last error: {last_error}</small>"
-        )
-
-
-# ─── Helpers ──────────────────────────────────────────────────────────────────
-def _extract_best(question: str, context: str) -> str:
-    """Keyword-based extractive fallback."""
-    keywords    = set(re.findall(r'\b\w{4,}\b', question.lower()))
-    best_chunk  = ""
-    best_score  = 0
-    for chunk in context.split("---"):
-        words = set(re.findall(r'\b\w{4,}\b', chunk.lower()))
-        score = len(keywords & words)
-        if score > best_score:
-            best_score = score
-            best_chunk = chunk.strip()
-    excerpt = best_chunk[:600] + ("..." if len(best_chunk) > 600 else "")
-    return excerpt or "No relevant content found."
-
-
-def _get_suffix(name: str) -> str:
-    return os.path.splitext(name)[-1].lower() or ".txt"
-ENDOFFILE
-echo "Done"
-Output
-
-Done
-
-Presented file
-Done
-Summary of what to do
-Do Step 1 & 2 first (get token + add secret) — that's the real fix.
-
-Then replace rag_engine.py with the new file above. The new version:
-
-Shows a clear friendly message if token is missing (instead of a cryptic error)
-Tries 4 different free models in order until one responds
-Falls back to extractive answer (pulls the most relevant paragraph from your doc) even if all LLMs fail
-After you add HF_TOKEN as a secret in Space Settings, the Space restarts automatically — no rebuild needed. Just ask a question again and it will work ✅
-
-Rag engine
-PY 
-
-
-
-
-
-
-Claude is AI and can make mistakes. Please double-check responses.
-Rag engine · PY
-Copy
-
-"""
-RAG Engine
-──────────
-- Embeddings : sentence-transformers/all-MiniLM-L6-v2  (HuggingFace, free)
-- Vector DB  : ChromaDB  (local, in-memory / persistent)
-- LLM        : HuggingFace Router API  (tries multiple free models)
-- Chunking   : Recursive character splitter with overlap
-"""
-
-import os
-import re
-import requests
-import tempfile
-from typing import Tuple, List
-
-import chromadb
-from chromadb.config import Settings
-from langchain.text_splitter import RecursiveCharacterTextSplitter
-from langchain_community.embeddings import HuggingFaceEmbeddings
-from langchain_community.vectorstores import Chroma
-from langchain_community.document_loaders import PyPDFLoader, TextLoader
-
-# ─── Configuration ─────────────────────────────────────────────────────────────
-EMBED_MODEL     = "sentence-transformers/all-MiniLM-L6-v2"
-CHUNK_SIZE      = 800
-CHUNK_OVERLAP   = 150
-TOP_K           = 4
-COLLECTION_NAME = "docmind_collection"
-CHROMA_DIR      = "./chroma_db"
-
-# Free models to try in order (all on HF router)
-CANDIDATE_MODELS = [
-    "mistralai/Mistral-7B-Instruct-v0.3",
-    "microsoft/Phi-3.5-mini-instruct",
-    "google/gemma-2-2b-it",
-    "HuggingFaceH4/zephyr-7b-beta",
-]
-
-
-class RAGEngine:
-    """Full RAG pipeline: ingest → embed → store → retrieve → generate."""
-
-    def __init__(self):
-        self._embeddings  = None
-        self._vectorstore = None
-        self._splitter    = RecursiveCharacterTextSplitter(
-            chunk_size=CHUNK_SIZE,
-            chunk_overlap=CHUNK_OVERLAP,
-            separators=["\n\n", "\n", ". ", " ", ""],
+        docs    = retriever.invoke(question)
+        context = "\n\n---\n\n".join(
+            "[Chunk {}]\n{}".format(i + 1, d.page_content) for i, d in enumerate(docs)
         )
-
-    # ── Lazy-load embeddings ───────────────────────────────────────────────────
-    @property
-    def embeddings(self):
-        if self._embeddings is None:
-            self._embeddings = HuggingFaceEmbeddings(
-                model_name=EMBED_MODEL,
-                model_kwargs={"device": "cpu"},
-                encode_kwargs={"normalize_embeddings": True},
-            )
-        return self._embeddings
-
-    # ── Ingest uploaded Streamlit file ────────────────────────────────────────
-    def ingest_file(self, uploaded_file) -> int:
-        suffix = _get_suffix(uploaded_file.name)
-        with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
-            tmp.write(uploaded_file.read())
-            tmp_path = tmp.name
-        return self.ingest_path(tmp_path, uploaded_file.name)
-
-    # ── Ingest from file path ─────────────────────────────────────────────────
-    def ingest_path(self, path: str, name: str = "") -> int:
-        suffix = _get_suffix(name or path)
-        loader = PyPDFLoader(path) if suffix == ".pdf" else TextLoader(path, encoding="utf-8")
-        raw_docs = loader.load()
-
-        for doc in raw_docs:
-            doc.metadata["source"] = name or os.path.basename(path)
-
-        chunks = self._splitter.split_documents(raw_docs)
-
-        self._vectorstore = Chroma.from_documents(
-            documents=chunks,
-            embedding=self.embeddings,
-            collection_name=COLLECTION_NAME,
-            persist_directory=CHROMA_DIR,
-            client_settings=Settings(anonymized_telemetry=False),
-        )
-        return len(chunks)
-
-    # ── Query ─────────────────────────────────────────────────────────────────
-    def query(self, question: str) -> Tuple[str, List[str]]:
-        if self._vectorstore is None:
-            return "⚠️ Please upload a document first.", []
-
-        retriever = self._vectorstore.as_retriever(
-            search_type="mmr",
-            search_kwargs={"k": TOP_K, "fetch_k": TOP_K * 3},
-        )
-        docs     = retriever.invoke(question)
-        context  = "\n\n---\n\n".join(f"[Chunk {i+1}]\n{d.page_content}" for i, d in enumerate(docs))
-        sources  = list({d.metadata.get("source", "Document") for d in docs})
-        answer   = self._generate(question, context)
+        sources = list({d.metadata.get("source", "Document") for d in docs})
+        answer  = self._generate(question, context)
         return answer, sources
 
-    # ── LLM: try each model until one works ───────────────────────────────────
     def _generate(self, question: str, context: str) -> str:
         hf_token = os.environ.get("HF_TOKEN", "")
 
         if not hf_token:
+            excerpt = extract_best(question, context)
             return (
-                "⚠️ **HF_TOKEN not set.**\n\n"
-                "To enable AI answers:\n"
+                "HF_TOKEN not set. To enable AI answers:\n"
                 "1. Get a free token at huggingface.co/settings/tokens\n"
-                "2. Add it as a **Secret** named `HF_TOKEN` in your Space Settings\n\n"
-                f"**Most relevant excerpt from your document:**\n\n{_extract_best(question, context)}"
+                "2. Add it as a Secret named HF_TOKEN in your Space Settings\n\n"
+                "Most relevant excerpt from your document:\n\n" + excerpt
             )
 
         headers = {
             "Content-Type":  "application/json",
-            "Authorization": f"Bearer {hf_token}",
+            "Authorization": "Bearer " + hf_token,
         }
         messages = [
             {
@@ -678,18 +128,27 @@ class RAGEngine:
             },
             {
                 "role": "user",
-                "content": f"Document context:\n{context}\n\nQuestion: {question}",
+                "content": "Document context:\n" + context + "\n\nQuestion: " + question,
             },
         ]
 
         last_error = ""
         for model_id in CANDIDATE_MODELS:
             try:
-                url  = f"https://router.huggingface.co/hf-inference/models/{model_id}/v1/chat/completions"
+                url  = (
+                    "https://router.huggingface.co/hf-inference/models/"
+                    + model_id
+                    + "/v1/chat/completions"
+                )
                 resp = requests.post(
                     url,
                     headers=headers,
-                    json={"model": model_id, "messages": messages, "max_tokens": 512, "temperature": 0.2},
+                    json={
+                        "model":       model_id,
+                        "messages":    messages,
+                        "max_tokens":  512,
+                        "temperature": 0.2,
+                    },
                     timeout=60,
                 )
                 if resp.status_code == 200:
@@ -697,26 +156,23 @@ class RAGEngine:
                     if answer:
                         return answer
                 else:
-                    last_error = f"{resp.status_code}: {resp.text[:200]}"
+                    last_error = str(resp.status_code) + ": " + resp.text[:200]
             except Exception as e:
                 last_error = str(e)
                 continue
 
-        # All models failed → extractive fallback
-        excerpt = _extract_best(question, context)
+        excerpt = extract_best(question, context)
         return (
-            f"*(All LLM models unavailable – showing most relevant excerpt)*\n\n"
-            f"{excerpt}\n\n"
-            f"<small>Last error: {last_error}</small>"
+            "LLM models unavailable - showing most relevant excerpt:\n\n"
+            + excerpt
+            + "\n\nLast error: " + last_error
         )
 
 
-# ─── Helpers ──────────────────────────────────────────────────────────────────
-def _extract_best(question: str, context: str) -> str:
-    """Keyword-based extractive fallback."""
-    keywords    = set(re.findall(r'\b\w{4,}\b', question.lower()))
-    best_chunk  = ""
-    best_score  = 0
+def extract_best(question: str, context: str) -> str:
+    keywords   = set(re.findall(r'\b\w{4,}\b', question.lower()))
+    best_chunk = ""
+    best_score = 0
     for chunk in context.split("---"):
         words = set(re.findall(r'\b\w{4,}\b', chunk.lower()))
         score = len(keywords & words)
@@ -727,6 +183,5 @@ def _extract_best(question: str, context: str) -> str:
     return excerpt or "No relevant content found."
 
 
-def _get_suffix(name: str) -> str:
+def get_suffix(name: str) -> str:
     return os.path.splitext(name)[-1].lower() or ".txt"
-