Apppublic

aryan-01001/knowledgebase

sourceHugging Faceupdated 4mo agoView on Hugging Face

0likes

App README

📁 Enterprise Document Search & Knowledge Base Workspace

A high-performance, citation-backed Retrieval-Augmented Generation (RAG) platform designed for engineering teams and internal documentation hubs. Inspired by the clean, information-dense visual languages of Notion, Linear, GitHub, and Confluence, this platform enables teams to index PDFs, run hybrid semantic + keyword searches across deep technical documentation, and extract citation-backed syntheses.

Clicking any citation immediately focuses the integrated PDF Modal Viewer exactly on the page where the source evidence was found.

🚀 Key Capabilities

—Hybrid Search (RRF): Merges dense semantic vectors (FAISS via fastembed) and sparse keyword matches (BM25 Okapi) using Reciprocal Rank Fusion (RRF) for optimal retrieval precision.
—Lightweight ONNX Embeddings: Uses fastembed to run all-MiniLM-L6-v2 locally via ONNX runtime — no PyTorch required, ~150MB RAM footprint.
—Real-Time Token Streaming: Answers stream token-by-token via Server-Sent Events (SSE) using Groq's streaming API.
—Precise Citation Backing: Every answer is strictly grounded on uploaded documents with page-level citation anchors ([1], [2], [3]).
—Disk Persistence: FAISS indices, BM25 registries, and chunk metadata persist to backend/data/indices/ and survive server restarts.
—Professional Workspace UI: Clean, tinted neutral palette modelled after Linear, GitHub, and Confluence — no AI chatbot aesthetics.

🏗️ System Architecture

mermaid

graph TD
    subgraph Ingestion ["1. Document Ingestion"]
        PDF[PDF Upload] --> Loader[PyPDFLoader]
        Loader --> Splitter[RecursiveCharacterTextSplitter\nchunk_size=800, overlap=200]
        Splitter --> Chunks[Text Chunks + Metadata]
        Chunks --> Embed[fastembed ONNX\nall-MiniLM-L6-v2]
        Embed --> FAISS[(FAISS Index\nPersisted to Disk)]
        Chunks --> BM25Build[BM25Okapi Indexer]
        BM25Build --> BM25[(BM25 Index\nPersisted to Disk)]
    end

    subgraph Retrieval ["2. Retrieval & Fusion"]
        Q[User Query] --> VecSearch[FAISS Vector Search\nTop-10]
        Q --> KwSearch[BM25 Keyword Search\nTop-10]
        FAISS -.-> VecSearch
        BM25 -.-> KwSearch
        VecSearch --> RRF[Reciprocal Rank Fusion\nk=60]
        KwSearch --> RRF
        RRF --> Top5[Top 5 Fused Chunks]
    end

    subgraph Generation ["3. Generation & Streaming"]
        Top5 --> Prompt[Context-Enriched Prompt]
        Q --> Prompt
        Prompt --> Groq[Groq SSE Stream\nllama-3.3-70b-versatile]
        Groq --> SSE[Server-Sent Events]
    end

    subgraph UI ["4. React Workspace Client"]
        SSE --> Stream[Typewriter Stream]
        Stream --> Citations[Click Citation]
        Citations --> PDF2[PDF Modal Viewer\nZoom + Paging]
    end

🛠️ Tech Stack

Backend

Component	Technology
Framework	`FastAPI` (Python 3.10+)
PDF Loading	`LangChain` — `PyPDFLoader` + `RecursiveCharacterTextSplitter`
Embeddings	`fastembed` — ONNX runtime, `all-MiniLM-L6-v2`, no PyTorch
Vector Store	`faiss-cpu` — `IndexFlatL2`, persisted to disk
Sparse Search	`rank-bm25` — BM25 Okapi, persisted to disk
Fusion	Reciprocal Rank Fusion (RRF, k=60)
Language Model	`Groq API` — `llama-3.3-70b-versatile`, SSE streaming

Frontend

Component	Technology
Framework	`React 19` + `Vite`
Styling	Vanilla CSS — Linear/Notion-inspired tinted neutrals
HTTP Client	`Axios` with upload progress hooks
PDF Rendering	`React-PDF` with web worker support
Streaming	Native `EventSource` / SSE

📂 Project Structure

Production-Rag/
├── backend/
│   ├── app/
│   │   ├── generator.py    # Groq LLM prompts & SSE streaming
│   │   ├── ingest.py       # PDF parsing, chunking, FAISS + BM25 persistence
│   │   ├── main.py         # FastAPI routes (upload, search, documents, delete)
│   │   ├── reranker.py     # RRF score-based sorting (no CrossEncoder)
│   │   └── retrieval.py    # fastembed ONNX embeddings, FAISS + BM25 + RRF
│   ├── data/
│   │   ├── documents/      # Uploaded PDFs served as static files
│   │   └── indices/        # FAISS .index + BM25 .pkl + chunks.json per doc
│   ├── .env                # GROQ_API_KEY
│   ├── Procfile            # Render start command
│   └── requirements.txt    # Minimal Python dependencies
├── frontend/
│   ├── src/
│   │   ├── App.jsx         # Main workspace panel + search console
│   │   ├── App.css         # Layout and card styles
│   │   ├── ChatMessage.jsx # Query / Synthesis timeline entries
│   │   ├── index.css       # CSS variables, color tokens, scrollbars
│   │   ├── PdfModal.jsx    # PDF viewer modal with zoom + paging
│   │   ├── Sidebar.jsx     # File explorer + upload zone
│   │   ├── SourcesPanel.jsx# Extracted evidence reference panel
│   │   └── Typewriter.jsx  # SSE token streaming printer
│   ├── .env                # VITE_API_URL (set to backend URL for prod)
│   └── vite.config.js
└── readme.md

⚙️ Running Locally

1. Backend

powershell

cd backend
python -m venv venv
.\venv\Scripts\Activate.ps1
pip install -r requirements.txt

Create backend/.env:

env

GROQ_API_KEY=your_groq_api_key_here

Start the server:

powershell

uvicorn app.main:app --reload

API runs at http://127.0.0.1:8000 — docs at http://127.0.0.1:8000/docs.

2. Frontend

powershell

cd frontend
npm install
npm run dev

Open http://localhost:5173.

For production, set in frontend/.env:

env

VITE_API_URL=https://your-backend.onrender.com

🚀 Production Deployment

Backend → Hugging Face Spaces (Docker Space)

1.Push repo to GitHub.
2.Go to Hugging Face Spaces and click Create new Space.
3.Settings:
4.Space name: your-rag-backend
5.License: mit (or your choice)
6.Select the Space SDK: Docker -> Blank
7.Space hardware: Free (Provides 16GB RAM and 2 vCPUs)
8.Under "Space settings", add your Repository Secret: GROQ_API_KEY
9.Since your code is on GitHub, you can link it directly or push your code to the Hugging Face git remote. Ensure the backend folder contains the Dockerfile.
10.Note: Because the Space starts from the root, ensure your Dockerfile is placed at the root or configure the Space to build from backend/Dockerfile. The provided Dockerfile assumes it's run from the backend directory.

(Alternative: You can simply copy the contents of your `backend` folder into the Hugging Face Space file editor if you don't want to use git).

Frontend → Vercel

1.Create a Project on Vercel:
2.Root Directory: frontend
3.Framework: Vite
4.Output: dist
5.Add Environment Variable: VITE_API_URL = your Render backend URL.
6.Deploy.