Coding Assessment Assistant

A retrieval-augmented assistant for coding assessments. It ingests DevDocs-style language documentation (and PDFs) into a local Qdrant vector store, retrieves relevant chunks with hybrid dense/sparse/late-interaction search, classifies user intent with a DistilBERT model to guard against off-topic or solution-seeking prompts (so it won't just hand over assessment solutions), and streams answers from a local Ollama model through a Streamlit chat UI.

Requirements

• Python 3.10+
• Ollama running locally with the following models pulled:

  ollama pull mistral
  ollama pull nomic-embed-text
  

Setup

git clone https://github.com/Koifish2004/coding-assessment-assistant.git
cd coding-assessment-assistant
python3 -m venv venv
source venv/bin/activate
pip install streamlit langchain-classic langchain-ollama langchain-core \
    qdrant-client fastembed transformers torch pymupdf4llm beautifulsoup4 \
    pandas requests

Add documents

• PDFs: drop files into data/*.pdf.
• Language docs: add a folder under data/<language>/ containing meta.json, db.json, and index.json (DevDocs export format).

Ingest

python -c "from ingestion.ingest import ingest_pdf, ingest; ingest_pdf(); ingest()"

This embeds and stores chunks in a local Qdrant instance at vectorstore/qdrant_db.

Run

streamlit run app/app.py

Open the URL Streamlit prints, pick a language, and start asking questions about your ingested documents.