Applied AI / Document retrieval / Project notes
SahayBot
Answers with a source. In two languages. A bilingual assistant that connects cooperative members to guidance grounded in reference PDFs.
The problem
Cooperative guidance is spread across lengthy reference documents. SahayBot gives users an English or Hindi interface to ask questions and inspect the original evidence.
How the system works
A React interface connects to FastAPI. PDFs become page-aware chunks with multilingual embeddings. Hybrid semantic and lexical retrieval finds evidence, and a local Ollama model produces a cited response.
A key engineering decision
SQLite stores document metadata, chunks, and normalized embedding vectors together. This makes the demo corpus portable and lets document replacements update their passages transactionally.
Implemented scope
PDF upload and replacement, source-page links, an educational EMI calculator, and editable grievance drafts. Drafts can be downloaded, while official grievance submission stays with the linked portals.
Scope & limitations
A working prototype for SIH 2026. Reference editions require source review. OCR, account isolation, physical kiosks, and grievance filing remain outside the implemented scope.
Next project