AI Research Assistant

A production-ready Retrieval-Augmented Generation (RAG) API that lets researchers query academic documents (text or PDF) in natural language. Sustains sub-200ms p95 latency under concurrent load with 95% retrieval consistency, backed by Pinecone vector search, LangChain orchestration, and OpenAI GPT-3.5-Turbo with a local Flan-T5 fallback.

---

Highlights

• Sub-200ms p95 latency under concurrent load
• 95% retrieval consistency across benchmark queries
• Automated regression testing via GitHub Actions
• Model fallback: OpenAI GPT-3.5-Turbo → local Flan-T5-Base when no API key
• PDF + text ingestion with automated conversion
• Streamlit UI for uploading, reindexing, querying, and summarization

---

Architecture

┌──────────────────┐    ┌──────────────────┐    ┌─────────────────┐
│  Streamlit UI /  │───▶│    LangChain     │───▶│    Pinecone     │
│    CLI Query     │    │  (Orchestration) │    │  (Vector Index) │
└──────────────────┘    └────────┬─────────┘    └────────┬────────┘
                                 │                       │
                                 │  Top-k passages       │
                                 ▼                       │
                        ┌──────────────────┐             │
                        │  Chunking Layer  │◀────────────┘
                        └────────┬─────────┘
                                 │
                                 ▼
                        ┌──────────────────┐
                        │  GPT-3.5-Turbo   │
                        │        or        │
                        │   Flan-T5-Base   │
                        │    (fallback)    │
                        └────────┬─────────┘
                                 │
                                 ▼
                        ┌──────────────────┐
                        │ Grounded Answer /│
                        │    Summary       │
                        └──────────────────┘

TODO: replace this ASCII sketch with a rendered diagram (Excalidraw or draw.io). Same shape, more polish.

---

Quick Start

git clone https://github.com/gauravch-code/AI-Research-Assistant.git
cd AI-Research-Assistant

python -m venv venv
source venv/bin/activate         # macOS / Linux
# venv\Scripts\activate           # Windows

pip install -r requirements.txt

Create a .env in the project root:

PINECONE_API_KEY=your-pinecone-key
PINECONE_ENVIRONMENT=your-pinecone-environment
OPENAI_API_KEY=sk-...             # Optional; omit to use local Flan-T5

Run the Streamlit app:

streamlit run frontend/streamlit_app.py

Or run a one-off query:

python run_query.py

---

How It Works

1. Embedding — Documents are embedded via HuggingFace all-MiniLM-L6-v2.
2. Indexing — Embeddings are upserted to Pinecone for approximate nearest-neighbor search.
3. Retrieval — Given a query, the top-k semantically relevant passages are fetched.
4. Chunking — Long passages are split to respect model context limits.
5. Generation — GPT-3.5-Turbo (or Flan-T5 fallback) produces a grounded answer or summary.

---

Project Structure

.
├── backend/
│   ├── rag_pipeline/rag_engine.py
│   ├── retriever/
│   │   ├── pinecone_setup.py
│   │   └── document_retriever.py
│   └── utils/
│       ├── document_loader.py
│       └── pdf_loader.py
├── frontend/streamlit_app.py
├── data/processed_docs/
├── run_query.py
├── requirements.txt
└── .env

---

Testing

Automated regression tests run on every push via GitHub Actions, covering:

• Embedding pipeline correctness
• Pinecone index consistency
• End-to-end query latency benchmarks

Run locally:

pytest tests/

---

Extension Points

• Swap model_name in rag_engine.py for GPT-4 or a local Llama model.
• Add hybrid or reranked retrieval (BM25 + dense).
• Fine-tune with LoRA / PEFT for domain-specific adaptation.

---

Contact

Gaurav Chintakunta · LinkedIn · Portfolio · gaurav.pvt25@gmail.com