Case Study
KnowledgeAssistant
Upload docs, search passages, get answers with source citations — no commercial vector DB required.
The Problem
Why this exists
Every team I've worked with has a "docs folder problem." SOPs, project notes, client records — they pile up and nobody can find anything. People end up asking whoever "might know," which is slow and breaks when that person is out.
How It Works
From input to outcome
Upload
Text files, markdown, or pasted notes. Each document gets tagged with metadata so you can track provenance.
Chunk and vectorize
Documents split into overlapping passages (~20% overlap so nothing gets cut off at boundaries). Each passage becomes a searchable vector embedding.
Ask and verify
Type a question in plain English. The system surfaces relevant passages with similarity scoring. Every answer includes a citation back to the source document — no black-box responses.
Technical Decisions
Why build it this way
No cloud vector database
In-process similarity search keeps deployment dead simple, cost at zero, and data fully private. For internal business docs, sending everything to Pinecone was a non-starter from day one.
Results
What it enables
New team members query internal docs immediately instead of waiting for walkthroughs
Search specific policies in seconds instead of digging through folders
Every answer is auditable — citations link back to the original document
What's Next
Planned improvements
- PDF/Word uploads with auto-extraction
- Batch upload folders
- Conversation history and query logs