PoliticGPT-Intern-Task

This repository contains my solution for the Kalpit Pvt Ltd AI Intern Hiring Assignment (Phase 1). It implements a local, command-line Q&A system using a Retrieval-Augmented Generation (RAG) pipeline. The system answers questions solely from the provided speech.txt (an excerpt from Dr. B. R. Ambedkar’s writing) — no internet or external APIs are used.

Problem Statement (paraphrased)

Build a simple CLI-based Q&A system that:

• Loads the provided speech.txt.
• Splits text into manageable chunks.
• Creates local embeddings and stores them in a local vector database.
• Retrieves relevant chunks for a user’s question.
• Generates an answer using the retrieved context and a local LLM.

Constraints and choices required by the brief:

• Language: Python 3.8+
• Orchestration: LangChain
• Vector store: ChromaDB (local)
• Embeddings: HuggingFace sentence-transformers/all-MiniLM-L6-v2 (local; no keys)
• LLM: Ollama with Mistral 7B (local; no accounts/keys)
• Deliverables: code (main.py), requirements.txt, README.md, and speech.txt in a public repo

Tech Stack

• Language: Python 3.8+
• Framework: LangChain (1.x module split)
• LLM: Ollama with Mistral 7B (local)
• Embeddings: sentence-transformers/all-MiniLM-L6-v2 (local)
• Vector store: ChromaDB (local)

How It Works (RAG Flow)

• Document loading: TextLoader reads speech.txt.
• Chunking: CharacterTextSplitter creates overlapping chunks.
• Embedding: HuggingFace model produces vector representations locally.
• Indexing: Chroma persists embeddings to ./chroma_db.
• Retrieval + Generation: Top matches are concatenated into context and passed to Ollama (Mistral) with a compact prompt to produce an answer.

Project Structure

• main.py — end-to-end RAG pipeline and CLI loop
• speech.txt — the source document used for answering
• requirements.txt — Python dependencies
• chroma_db/ — persisted local vector store (auto-created)

Setup

Below are Windows-focused steps (PowerShell).

1) Install Ollama and pull the model

• Windows: download and install from https://ollama.ai/ (launch the app so the service runs)
• Pull the model:

ollama pull mistral

Optional, to run the server explicitly:

ollama serve

2) Python environment and dependencies (Windows PowerShell)

Use the bundled virtual environment (gptintern) or create your own.

• Activate the existing venv:

& "D:Path/Ambedkar_GPT Intern Task/gptintern/Scripts/Activate.ps1"

• Or create a fresh one:

python -m venv .venv
& ".\.venv\Scripts\Activate.ps1"

• Install dependencies:

python -m pip install -r requirements.txt

Run

With your environment active and Ollama running:

python main.py

Then type your question at the prompt, or type exit to quit.

Troubleshooting

• ModuleNotFoundError for langchain.document_loaders or langchain.text_splitters:
  • LangChain 1.x split modules. This project uses:
    • langchain_community.document_loaders
    • langchain_text_splitters
    • langchain_community.embeddings / vectorstores / llms
  • Ensure dependencies are installed:

python -m pip install langchain langchain-community langchain-text-splitters chromadb sentence-transformers

• Ollama connection issues:

  • Make sure the service is running and mistral is pulled.
  • Try ollama run mistral once to confirm the model executes.

• First run is slow:

  • Embedding the corpus and creating the Chroma index happens only once and is cached under ./chroma_db.

Notes on Implementation

• The classic RetrievalQA chain referenced in the brief isn’t present in recent LangChain 1.x. This repo composes the chain using langchain_core runnables (ChatPromptTemplate → Ollama → StrOutputParser) with a retriever from Chroma.
• Everything runs fully offline once models are pulled (Ollama + HF embeddings).

Deliverables (as required)

• Public repository with:
  • main.py (well-commented)
  • requirements.txt
  • README.md (this file)
  • speech.txt

License

For hiring evaluation purposes only.