title	AdaptiveRAG
sdk	docker
pinned	true
license	mit
short_description	Agentic + Self-RAG + Modular RAG with visual pipeline UI

📚 AdaptiveRAG

Production-grade RAG combining Modular · Self-RAG · Agentic patterns

Every stage of the pipeline is visualized live — from raw text to grounded answer with citations.

🚀 Try Live Demo · 💻 Run Locally

---

🎬 Demo

📹 Watch full pipeline demo →

Shows: question encoding → Self-RAG routing → hybrid retrieval → 2D vector space → self-critique

---

🧠 What makes this different

Most RAG demos do: embed query → cosine search → stuff into prompt. This does:

User question
   ↓  embed (MiniLM-L6 → 384-dim vector)
   ↓  Self-RAG router  →  RETRIEVE / ANSWER_DIRECTLY / CLARIFY
   ↓  Planner          →  break into focused sub-queries
   ↓  Dense retrieval  →  ChromaDB cosine similarity (k=12)
   ↓  Sparse retrieval →  BM25 keyword matching (k=12)
   ↓  RRF fusion       →  Reciprocal Rank Fusion merge
   ↓  Cross-encoder    →  BGE reranker deep relevance scoring (top 5)
   ↓  LLM answer       →  Qwen3-VL (local) / LLaMA 3.1 via Groq (hosted)
   ↓  Self-critique     →  grounded? complete? confidence score
   ↓  Refine & retry   →  if confidence < 0.85
   →  Answer + citations + trace

---

🔬 Underhood Pipeline View

Every step renders its inputs and outputs as it runs:

Step	What you see
1 · Question encoding	Embedding model · 384 dimensions · L2 norm · latency · first-32-dim bar chart · raw vector[0:8] values
2 · Self-RAG router	Color-coded decision pill (RETRIEVE / ANSWER_DIRECTLY / CLARIFY) + LLM reasoning
3 · Planner	Sub-query cards with rationale for each step
4 · Dense retrieval	Cosine similarity bar chart + chunk cards with scores
4 · Sparse retrieval	BM25 normalized score chart + chunk cards
4 · RRF fusion	Merged ranking chart showing how both lists combine
4 · Cross-encoder rerank	BGE relevance score chart (final top-5)
4 · Vector space	2D PCA scatter — query vs all hits, colored by source (dense / sparse / both)
5 · Context assembly	Exact passages handed to the LLM, with metadata
6 · Self-critique	Grounded ✅ · Complete ✅ · Confidence score vs threshold

---

🗂️ Knowledge Base

14 foundational AI papers pre-indexed as 1,934 semantic chunks:

Category	Papers
Transformers	Attention Is All You Need · BERT · GPT-3
Diffusion	DDPM · DDIM
RAG	RAG Original · RAG Survey · Self-RAG · HyDE
Vision	ViT · CLIP
Agents	ReAct · Chain-of-Thought
LLMs	LLM Survey

---

⚙️ Tech Stack

Component	Tool	Why
Vector DB	ChromaDB (local)	No API cost, persistent
Dense embeddings	all-MiniLM-L6-v2	Fast, 384-dim, normalized
Sparse retrieval	rank-bm25 (BM25Okapi)	Keyword precision
Fusion	Reciprocal Rank Fusion	Combines rankings without score normalization
Reranker	BAAI/bge-reranker-base	Cross-encoder, deep relevance scoring
LLM (local)	Qwen3-VL 8B via Ollama	Vision-language, runs on Apple Silicon
LLM (hosted)	LLaMA 3.1 8B via Groq	Free API, fast inference
UI	Streamlit	Fast to build, easy to demo

---

🚀 Run Locally

git clone https://github.com/Gh-Novel/AdaptiveRAG
cd AdaptiveRAG

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# start Ollama with the vision-language model
ollama serve
ollama pull qwen3-vl:8b-instruct-q8_0-optimized

streamlit run app.py

Or use the CLI:

.venv/bin/python ask.py "How does Self-RAG decide when to retrieve?"

---

☁️ Hosted on Hugging Face

The live demo runs on HF Spaces (CPU free tier) with Groq API handling LLM calls.

• Embedding + retrieval + reranking run locally inside the container (MiniLM + BGE)
• GROQ_API_KEY secret drives routing, planning, answering, and self-critique
• Pre-built index (1,934 chunks, ~59 MB) is committed via git-lfs — no ingestion on startup

🤗 Open Live Demo