AI Documentation Bot

An intelligent RAG bot that answers natural-language questions across Pandas, NumPy, and Scikit-learn documentation, using an MLP classifier to rank retrieved snippets by source reliability. Answer relevance improved by 20% by prioritizing official API references over community content like Stack Overflow.

---

Highlights

• 20% relevance improvement by prioritizing official docs over community answers
• 3 large technical corpora indexed (Pandas, NumPy, Scikit-learn)
• MLP-based source classifier trained on TF-IDF features to score API-spec vs tutorial vs community
• LLM flexibility: OpenAI (GPT-4o / GPT-3.5-Turbo) or local Ollama (Llama 3.1)
• Streamlit UI for quick prototyping

---

Architecture

┌──────────────┐   ┌──────────────┐   ┌────────────────┐
│   Scraper    │──▶│  corpus.json │──▶│ Sentence-BERT  │
│(BeautifulSoup)│   │              │   │   Embeddings   │
└──────────────┘   └──────────────┘   └────────┬───────┘
                                               │
                                               ▼
                                       ┌──────────────┐
                                       │   Pinecone   │
                                       │ Vector Store │
                                       └──────┬───────┘
                                              │
User Query ──▶ Streamlit ──▶ Retrieval ───────┘
                                │
                                ▼
                     ┌────────────────────┐
                     │  MLP Source        │  ← ranks snippets by
                     │  Classifier        │    API-spec / tutorial /
                     │  (TF-IDF + MLP)    │    community reliability
                     └──────────┬─────────┘
                                ▼
                     ┌────────────────────┐
                     │ Prompt Assembler   │  ← prioritizes official
                     │ (LangChain)        │    sources in context
                     └──────────┬─────────┘
                                ▼
                     ┌────────────────────┐
                     │  LLM (OpenAI /     │
                     │  local Ollama)     │
                     └──────────┬─────────┘
                                ▼
                          Grounded Answer

TODO: replace with a rendered diagram once you have time. This sketch shows the flow.

---

Tech Stack

Backend & ML   Python 3.9+ · Scikit-learn (MLPClassifier, TF-IDF) · Sentence Transformers
Vector DB   Pinecone
LLM   OpenAI (GPT-4o / GPT-3.5-Turbo) or local Ollama (Llama 3.1 8B)
Orchestration   LangChain
Scraping   Requests · BeautifulSoup4
UI   Streamlit

---

Quick Start

git clone https://github.com/gauravch-code/AI-doc-bot.git
cd AI-doc-bot

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

pip install -r requirements.txt

Create a .env in the root:

PINECONE_API_KEY=your-pinecone-key
OPENAI_API_KEY=your-openai-key    # Omit if using Ollama exclusively

First-time setup (three prep steps)

python scraper.py             # collect docs → corpus.json
python train_classifier.py    # train MLP classifier → .pkl files
python build_vectordb.py      # populate Pinecone index

Run the app

streamlit run streamlit_app.py

Opens at http://localhost:8501.

Optional: use a local LLM via Ollama

# install Ollama from https://ollama.com/download
ollama pull llama3.1:8b
# swap ChatOpenAI for ChatOllama inside streamlit_app.py

---

Testing

A test suite covering scraper HTML parsing, mocked Pinecone/OpenAI responses, and classifier loading is in progress. Run manual smoke tests via the Streamlit UI across a mix of Pandas, NumPy, and Scikit-learn queries to verify citation quality and source prioritization.

---

Roadmap

• Extend to Matplotlib, TensorFlow, PyTorch documentation
• Query-time library detection (auto-route to the right index namespace)
• Replace TF-IDF with sentence embeddings in the classifier
• Dockerize and deploy to Streamlit Community Cloud
• Add formal unit and integration tests with mocking

---

Contact

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