SLM Profile RAG Chatbot

Overview

A Python-native RAG chatbot for professional profiles. It ingests resumes, project writeups, and portfolio docs, embeds them into ChromaDB, and serves grounded answers locally through Ollama with a Streamlit UI so recruiters can “interview” your work—not just skim a PDF.

Motivation

Job interviews can feel like a performance tax, especially for introverts who know the work but stumble under pressure. I built this in a weekend so quieter developers can let their work speak for itself. The stack is transparent and configurable: fork it, drop in your docs, tune YAML, and deploy a shareable chatbot (HF Spaces or local).

Tech Stack & Architecture

Architecture overview of the SLM Profile RAG stack

• UI: Streamlit chatbot, shareable link for recruiters/hiring managers.
• RAG Pipeline (LangChain): RetrievalQA chain stitching prompt, retriever, and LLM.
• Vector Store: ChromaDB (persistent) + all-MiniLM embeddings; similarity/MMR search.
• LLM Serving: Ollama with small, fast models (llama3.2:3b recommended; phi3:mini, gemma2:2b; llama3.1:8b with GPU).
• Config-First: YAML persona/model/chunking; .env overrides for URLs/paths; no hard-coding.
• Packaging: UV + Hatchling + Versioningit; Ruff for lint/format; Docker image for HF Spaces.
• Document Processing: PDF/DOCX/HTML/MD/TXT loaders with configurable chunk size/overlap; optional prioritized main document for guaranteed context.
• Response Polishing: Optional enhancer for concise, recruiter-friendly tone with citations.

Workflow

• Indexing: python -m src.build_vectorstore to load, chunk, embed, and persist documents in ChromaDB.
• Query: Embed question, retrieve top-k chunks, compose prompt, and answer via Ollama with citations in Streamlit.
• Config Tweaks: Swap models, adjust chunking, or change persona via config.yaml and .env.

Deployment and Packaging

• Local: Streamlit + Ollama + local ChromaDB.
• Hugging Face Spaces: Docker-based Space installs Ollama, pulls model, restores persistent chroma_db, and serves the app over HTTPS.
• Build System: UV + Hatchling with Versioningit for git-tag-based versions; Ruff for lint/format.
• Config: .env overrides config.yaml (model, chunking, paths).

Why It Matters

• Lets your work speak when interviews get noisy—grounded answers from your own corpus.
• Privacy-first: runs locally by default; no external API calls required.
• Portable and shareable: one-click HF Space or run on a laptop with small models.
• Transparent and teachable: clean Python stack to learn RAG end-to-end.

Example UI

Streamlit chatbot interface showing the profile-aware assistant