Cytron HR Assistive Intelligence
Tech Specs
A chat assistant that answers employee questions only from official HR documents, cites every statement, and says so when it doesn't know.
Retrieval
- Vector index
- ChromaDB, one collection per embedding model
- Chunking
- Split by document structure; list items and tables are never cut; short sections merged
- Re-indexing
- Content-hash chunk ids, so only changed text is embedded again
- Relevance gate
- 0.60 similarity cut-off plus a 0.15 retrieval margin, both adjustable
- Calibration
- For gemini-embedding-2: on-topic questions score above 0.70; off-topic fall to 0.48–0.59
Answers
- Providers
- Google Gemini
- OpenAI
- Interchangeable for
- Both embeddings and answers
- Delivery
- Streamed token by token over server-sent events, with fallback models
- Citations
- Numbered, and only the cited excerpts are returned as sources
- Out-of-scope questions
- Fixed “not found” reply that points to HR; 0 model calls
Documents
- Formats
- DOCX
- TXT
- Markdown
- Maximum size
- 20 MB per file
- DOCX parsing
- Native OpenXML reader that includes floating text boxes and tables
Privacy & Quality
- Private chats
- No messages stored
- Usage analytics
- Never record question text
- Evaluation
- Saved test questions score retrieval after every change
- Insights
- Unanswered questions and answers rated down
Software
- Backend
- Python
- FastAPI
- SQLite
- Serving
- Single process for the API and the compiled React interface
- Interface
- React 18
- Vite
- Tailwind CSS
- Markdown rendering
- Citation viewer