Project · WRK-009
Internal Knowledge Base Assistant
An internal Q&A tool for an equipment after-sales team. Every answer comes with the source passage, so people can check it.
- Python
- FastAPI
- pgvector
- React
- LLM API
Project background
The after-sales team has hundreds of manuals, warranty policies and FAQ documents. New staff either dig through folders or interrupt a senior colleague.
Problem
- Documents are scattered across shared drives and chat history, in different versions;
- New hires ramp up slowly, and experienced staff get interrupted constantly;
- A general-purpose AI chatbot isn't trustworthy when you can't see where the answer came from.
Approach
Answer only from the company's own documents, and attach the quoted passage and page number to every answer. If there's no supporting source, say so.
Key features
- Document library by department and product, chunked and indexed on upload;
- Answers with numbered citations that open the original text;
- Permissions so each department only searches what it can see;
- Thumbs-down feedback, reviewed by an admin weekly;
- Old versions retire automatically when a document is replaced.
Architecture
FastAPI back end, vector search with PostgreSQL's pgvector extension, React front end. The language model is called through an API and can be swapped to suit the client.
What we handled
Solution design, the document pipeline, retrieval and answering, permissions, and building the test set.
How it went
With the after-sales team, we collected 120 real questions as a test set. Every change to retrieval gets run against it: how many are right, and are the citations accurate?
Screenshots
Test environment with sample documents.
What went wrong
Manuals are full of tables, and chunking kept separating rows from their headers. Tables are now handled separately and returned with their headers.
Outcome
In internal trial. Most test-set questions come back with the correct citation. Still tuning.
Current status
In development. A wider rollout to more departments is planned for next month.
Next
Connect it to the ticketing system so staff can search while handling a ticket.