Agentic-RAG Researcher

Hybrid Context Question Answering System with Web Search and Citation-Aware Generation

A production-oriented Agentic RAG pipeline designed for researchers and engineers to query academic literature using a combination of:

• Local research paper retrieval
• Real-time web search
• Intelligent query routing
• Reranking
• Citation-grounded generation
• Automated evaluation

The system dynamically decides whether a query should be answered using:

• Internal research documents
• External web knowledge
• A hybrid combination of both

---

Architecture Overview

flowchart TD

A[User Query]

A --> B[Intent Router]

B -->|Local| C[Vector Retrieval]
B -->|Web| D[Web Search]
B -->|Hybrid| C
B -->|Hybrid| D


C --> E[Qdrant Vector Database]

D --> F[Tavily / Brave Search]


E --> G[Reranker]

F --> G


G --> H[Context Builder]


H --> I[LLM Generation]


I --> J[Answer + Citations]

J --> K[Streamlit UI]

Loading

---

Key Features

Intelligent Query Routing

The system uses an LLM-based router to classify queries:

Question
   |
   ▼
Router

   |
   ├── local
   |
   ├── web
   |
   └── hybrid

Examples:

Query	Route
Explain transformer architecture from papers	Local
Latest AI announcements	Web
Compare research papers with current industry trends	Hybrid

---

Document Intelligence Pipeline

Research papers are processed through:

PDF

 |
 ▼

Layout-aware Parser

 |
 ▼

Hierarchical Chunking

 |
 ▼

Metadata Extraction

 |
 ▼

Embedding Generation

 |
 ▼

Qdrant Index

The ingestion pipeline preserves:

• Paper title
• Authors
• Page numbers
• DOI information
• Source metadata

---

Retrieval Architecture

The retrieval pipeline contains:

Semantic Retrieval

Vector similarity search using:

• Qdrant
• Dense embeddings

Reranking

Retrieved documents are refined using:

• Cross encoder reranking
• Relevance filtering

Pipeline:

Query

 |
 ▼

Vector Search

 |
 ▼

Top-K Documents

 |
 ▼

Reranker

 |
 ▼

LLM Context

---

Agent Workflow

The system is orchestrated using LangGraph.

Graph:

                 Query

                   |

                   ▼

             Intent Router

          /       |       \

         /        |        \

      Local      Web     Hybrid

        |          |        |

        ▼          ▼        ▼

     Qdrant     Search   Combined

          \       |       /

              Generation

                  |

                  ▼

          Citation Answer

---

Tech Stack

Backend

• Python
• FastAPI
• LangGraph
• Async architecture

LLM

• OpenAI API
• Embedding models

Vector Database

• Qdrant

Retrieval

• Semantic search
• Hybrid retrieval
• Cross encoder reranking

Web Search

• Tavily API
• Brave Search API

Frontend

• Streamlit

Evaluation

• Ragas

Infrastructure

• Docker
• Docker Compose
• GitHub Actions

---

Project Structure

agentic-rag-researcher/
├── src/
│   ├── api/
│   ├── graph/
│   ├── ingestion/
│   ├── retrieval/
│   ├── search/
│   ├── llm/
│   ├── evaluation/
│   └── utils/
├── scripts/
├── tests/
├── streamlit/
├── docker/
├── data/
├── docker-compose.yml
├── pyproject.toml
└── README.md

---

Running Locally

1. Clone Repository

git clone https://github.com/<username>/agentic-rag-researcher.git

cd agentic-rag-researcher

---

2. Create Environment

python -m venv .venv

source .venv/bin/activate

Windows:

.venv\Scripts\activate

---

3. Install Dependencies

pip install -e .

---

4. Configure Environment Variables

Create:

.env

Example:

OPENAI_API_KEY=

QDRANT_URL=http://localhost:6333

TAVILY_API_KEY=

BRAVE_API_KEY=

---

Start Infrastructure

Run:

docker compose up

Services:

FastAPI
    |
    localhost:8000


Qdrant
    |
    localhost:6333


Streamlit
    |
    localhost:8501

---

Ingest Research Papers

Place PDFs:

data/papers/

Run:

python scripts/ingest_papers.py

---

Start Application

Backend:

uvicorn src.main:app --reload

Frontend:

streamlit run streamlit/app.py

---

API Usage

Endpoint:

POST /api/v1/query

Request:

{
    "query":
    "Explain retrieval augmented generation"
}

Response:

{
    "query":
    "Explain retrieval augmented generation",

    "route_taken":
    "local",

    "answer":
    "RAG combines retrieval systems with language models..."
}

---

Evaluation

The project includes automated RAG evaluation.

Metrics:

Faithfulness

Measures whether answers are supported by retrieved evidence.

Answer Relevance

Measures whether the response answers the user's question.

Context Precision

Measures retrieval quality.

Run:

python scripts/evaluate_rag.py

---

Engineering Highlights

This project demonstrates:

✅ Agentic workflow orchestration
✅ Stateful graph execution
✅ Hybrid retrieval architecture
✅ Vector database management
✅ Async API design
✅ Citation-aware generation
✅ RAG evaluation pipeline
✅ Production-style repository organization

AI-Assisted Development

This project was developed with the assistance of AI coding tools for code review, refactoring suggestions, documentation support, and debugging assistance. The architecture, engineering decisions, and final implementation were designed and validated by the developer.

---

Future Improvements

Planned:

• Streaming token responses
• Authentication layer
• Distributed vector indexing
• Better document parsing with LlamaParse
• Human feedback loop
• Retrieval analytics dashboard

---

License

MIT License