Closed-Loop Adaptive RAG

A Retrieval-Augmented Generation (RAG) system that improves response reliability using a semantic faithfulness evaluation loop.

Unlike conventional RAG systems that generate responses after a single retrieval step, this implementation continuously evaluates the generated answer against the retrieved context. If the answer does not satisfy a semantic similarity threshold, the system automatically expands the retrieval context and regenerates the response.

This reduces hallucinations and increases factual consistency.

---

Features

• Retrieval-Augmented Generation (RAG)
• Semantic Vector Search
• Automatic Faithfulness Evaluation
• Closed Feedback Loop
• Dynamic Context Expansion
• Gemini-based Answer Generation
• Retry Mechanism for Low-Confidence Responses

---

Workflow

User Query ↓ Retrieve Top-K Chunks ↓ Generate Draft Answer ↓ Semantic Faithfulness Evaluation ↓ Faithfulness ≥ Threshold? ├── Yes → Return Answer └── No ↓ Increase Retrieval Size ↓ Retrieve Again ↓ Generate Again

---

Project Architecture

User Query
        │
        ▼
Vector Database
        │
        ▼
Relevant Context
        │
        ▼
LLM (Gemini)
        │
        ▼
Draft Answer
        │
        ▼
Semantic Similarity Checker
        │
 ┌──────┴────────┐
 │               │
Pass          Fail
 │               │
 ▼               ▼
Answer     Expand Retrieval
                 │
                 ▼
            Generate Again

---

Repository Structure

modules/
    retrieval/
        vector_db.py
    generation/
        llm_client.py
    evaluation/
        faithfulness.py

main.py

---

Technologies Used

• Python
• Sentence Transformers
• NumPy
• Google Gemini API
• FAISS / Vector Embeddings
• PDF Processing
• Cosine Similarity

---

Installation

Clone the repository

git clone https://github.com/username/closed-loop-adaptive-rag.git

Install dependencies

pip install -r requirements.txt

Run

python main.py

---

Novelty

This project introduces a Closed-Loop Adaptive Retrieval mechanism.

Traditional RAG

Retrieve
↓

Generate
↓

Answer

Adaptive RAG

Retrieve
↓

Generate
↓

Evaluate Faithfulness
↓

Low Score?

↓

Expand Retrieval

↓

Generate Again

The iterative retrieval strategy allows the system to dynamically increase contextual information whenever hallucination risk is detected.

---

Future Work

• Adaptive chunk sizing
• Multi-document retrieval
• Cross-encoder reranking
• Confidence estimation
• Knowledge graph integration

---

License

MIT License