Internal demo
Docsense
Answers staff questions from internal policy documents and cites the paragraph it used.
- Type
- AI
- Year
- 2026
- Stack
- Python, LangGraph, pgvector, OpenAI API, Next.js
Interface shown with sample data.
The challenge
HR and operations teams answer the same policy questions every week. The answers exist, spread across PDFs and wiki pages nobody can search well.
Who it’s for
Employees asking questions; HR owners maintaining the documents.
Solution and product strategy
- Documents are split by heading, embedded and stored with their source location.
- Answers quote the source paragraph and link to it; if nothing relevant is found, the assistant says so.
- Document owners see which questions went unanswered so they can fill the gaps.
System architecture
The path a request takes through the system, and what each part is responsible for.
- 01IngestPDF and wiki connectors, chunked by structure
- 02Vector storepgvector alongside metadata and permissions
- 03Answer serviceRetrieval, re-ranking and a cited answer
- 04Eval suite120 question/answer pairs scored on each release
What we’d measure
This is an internal demo, so there are no client results to report. These are the numbers a real deployment would be judged on.
- Answer accuracy on the evaluation set
- Share of questions answered with a citation
- Repeat questions reaching HR
What we learned
- Chunking by document structure beat fixed-size chunks on every test.
- "I couldn't find that" is a feature, not a failure.
Want something like this for your team?
Tell us what you're trying to do. A senior engineer reads every message and replies personally.