Spaces:
Sleeping
Sleeping
metadata
title: IPARD III RAG
emoji: 🌾
colorFrom: green
colorTo: blue
sdk: docker
pinned: false
app_port: 7860
IPARD III Document Assistant 🤖🌾
A production-grade Retrieval-Augmented Generation (RAG) system built specifically for IPARD III program documents. Using hybrid search, contextual retrieval, and intelligent query routing, it delivers accurate, context-aware answers about IPARD III grants, measures, and application processes.
✨ Features
Retrieval
- Hybrid Search: BM25 keyword search + Turkish-E5-Large semantic embeddings
- Reciprocal Rank Fusion (RRF): Intelligently merges keyword and semantic results
- Cross-Encoder Reranking: BGE-reranker-base for precision scoring (0.95+ scores)
- Neighboring Chunks: Previous and next chunks are included alongside retrieved chunks for fuller context
- Contextual Retrieval (Anthropic method): Each chunk is enriched with a 2-3 sentence LLM-generated context before embedding — reduces retrieval failures by ~67%
Routing & Intelligence
- Query Router: Llama-3.1-8b-instant automatically detects the relevant measure and document type from the user's question — even when no filter is selected
- Identity Bypass: Meta questions like "Who are you?" skip the RAG pipeline and are answered directly from the system prompt
- Metadata Filtering: Filter by measure (101/103/201/202/302) and document type
System
- Real-time Streaming: Token-by-token responses via Server-Sent Events (SSE)
- Dynamic System Prompt: Update LLM behavior via
system_prompt.txtwithout touching code - Router Prompt: Full measure/sector/sub-sector hierarchy defined in
router_prompt.txt - Turkish-Optimized: Turkish-E5-Large asymmetric model with passage/query prefix support
🏗️ Architecture
User Question
↓
┌─────────────────────────────┐
│ Query Router │
│ (Llama-3.1-8b-instant) │
│ measure + doc_type detect │
└─────────────┬───────────────┘
↓
┌─────────────────────────────┐
│ Hybrid Retrieval │
│ BM25 (20) + Semantic (20) │
│ → RRF Merge │
│ → Cross-Encoder Rerank (5) │
│ + Neighboring Chunks │
└─────────────┬───────────────┘
↓
┌─────────────────────────────┐
│ LLM Generation │
│ GPT-OSS 120B (Groq API) │
│ + Dynamic System Prompt │
└─────────────────────────────┘
🚀 Technology Stack
| Component | Technology | Purpose |
|---|---|---|
| Frontend | Streamlit | Web interface |
| Backend API | FastAPI | REST endpoints & streaming |
| Vector Database | ChromaDB | Embedding storage & retrieval |
| Keyword Search | BM25 (rank-bm25) | Lexical search |
| Embeddings | Turkish-E5-Large | Semantic text representation |
| Reranking | BGE-reranker-base | Precision improvement |
| Query Router | Llama-3.1-8b-instant (Groq) | Automatic measure detection |
| LLM | GPT-OSS 120B (Groq) | Answer generation |
| Deployment | Docker + HuggingFace Spaces | Containerized hosting |
📊 Performance
- Documents: 75+ Turkish IPARD III PDFs
- Chunks: 7,717 segments
- Embedding Dimension: 1,024
- Retrieval Pipeline: 20 BM25 + 20 Semantic → RRF → 5 Reranked + Neighboring Chunks
- Contextual Enrichment: LLM-generated context prepended to each chunk via DeepSeek API
- Response Time: ~2-5 seconds
🛠️ Setup
Requirements
- Python 3.11+
- Groq API key
Local Setup
git clone https://github.com/denizkaya20/IPARD-AI-Assistant.git
cd IPARD-AI-Assistant
pip install -r requirements.txt
echo "GROQ_API_KEY=your_key_here" > .env
# Backend
uvicorn src.api:app --reload --port 8000
# Frontend
streamlit run src/app.py --server.port 7860
Docker
docker build -t ipard-rag .
docker run -p 7860:7860 -e GROQ_API_KEY=your_key_here ipard-rag
🌐 HuggingFace Spaces Configuration
Required Secrets
| Secret | Description |
|---|---|
GROQ_API_KEY |
Groq API key (used for LLM + Query Router) |
Key Files
| File | Description |
|---|---|
src/system_prompt.txt |
LLM identity, measure info, response rules |
src/router_prompt.txt |
Full measure/sector/sub-sector hierarchy for routing |
data/all_chunks.json |
Contextually enriched document chunks |
data/embeddings.npy |
Pre-computed embedding vectors |
⚠️ Disclaimer
This is an independent project with no affiliation to TKDK or any official institution. Responses are AI-generated and do not constitute official guidance. Always refer to tkdk.gov.tr for authoritative information.
🙏 Acknowledgments
- TKDK — IPARD III documentation
- Groq — LLM API
- Hugging Face — Model hosting and Spaces
- Anthropic — Contextual Retrieval method