All selected work

Local AI / Retrieval systems / Project notes

Salaahkar

Study with your sources close by. A local RAG study assistant for PDFs, handwritten notes, and lecture transcripts, built around hybrid retrieval and cited answers.

PythonLangChainChromaDBOllama
SalaahkarSystem overview
01PDF / OCR / transcript
02Vector + BM25 search
03Ollama inference
04Answer + sources
Local inference · hybrid retrievalArchitecture diagram

The problem

University study material lives across textbooks, handwritten notes, and recorded lectures. I built Salaahkar to bring those sources into one searchable study environment for the AIML curriculum.

How the system works

PDF text, Tesseract OCR output, and YouTube transcripts are indexed in ChromaDB. Semantic vector search and BM25 keyword retrieval work together, with multi-query expansion and local Ollama inference.

A key engineering decision

Combining vector retrieval with keyword search helps cover both conceptual questions and precise terms. The Streamlit interface exposes retrieval depth and model temperature so the study workflow can be adjusted.

Implemented scope

Hybrid retrieval, multi-query expansion, OCR ingestion, transcript ingestion, and expandable source citations are implemented. The interface provides session-level model and retrieval controls.

Scope & limitations

SQLite chat history, subject-specific filtering, specialized mathematical parsing, Anki export, and study-guide generation are planned extensions. YouTube transcript ingestion requires access to the source service.

Next project

SahayBot

Explore case study