Trust
Cited, grounded answers
Each statement links to the document section it came from, and a click opens the excerpt so the reader can verify it.
Cytron HR Assistive Intelligence
Answers it can prove.
A chat assistant that answers employee questions only from official HR documents, cites every statement, and says so when it doesn't know.
Answered from 2 sources
Every question to one person.Employees ask. The documents answer.
Cytron has a single HR person, and each leave, claim or policy question pulled them away from real HR work. Now employees ask in plain language and get a cited answer straight from the official documents.
How does this policy apply to me?
Answered only from the official HR documents.1
No message to HR. No waiting for a reply.
Policies change.Change a document, not the system.
HR adds or removes documents in a knowledge base manager, and only the text that changed is embedded again, so the assistant stays current without a rebuild.
Only the changed chunk is embedded again.
A wrong answer costs trust.Accurate, or honest about it.
A relevance gate checks every question against the documents. If nothing is close enough, a fixed reply points the employee to HR and the language model is never called. Every statement carries a citation that opens its source.
On-topic questions
above 0.70, answered with citations
Off-topic questions
0.48 – 0.59, “not found” reply, model never called
Tuning by eye.Measure every change.
Each test case is a question plus the document that should answer it, or a flag that it must be refused. After any change to chunking, thresholds or models, a run scores retrieval against them without calling the answer model, and can suggest the best cut-off.
- Answered from the right documentShare of answerable questions where the expected document passed the cut-offs
- Mean reciprocal rankHow high the first correct passage ranks, where 1.0 means always first
- Off-topic correctly refusedShare of must-refuse questions where nothing passed, so the model is never called
- Every run is saved with a snapshot of its settings, so each change shows what it did.
See it in action.
A plain-language question, an answer built from the official documents, and numbered sources underneath.
Privacy
Private where it counts.
Policies and contracts don’t belong in a public chat. Here the index is stored locally, chats can leave no trace, and analytics never record what was asked. The AI provider is your choice.1
A local index
Documents are split and stored in a local ChromaDB index, with a separate collection for each embedding model.
Private chats
A private chat stores no messages at all.
Analytics without questions
Usage analytics keep only the outcome, timing and cited document names, never what was asked.
Get the highlights.
By design
Built to guarantee, not to guess.
The assistant is in internal pilot, so these are properties of the design rather than adoption figures.
- Model calls when nothing matches
- 0
- Similarity cut-off
- 0.60
- Interchangeable AI providers
- 2
- Supported file types
- 4
Out-of-scope questions are refused before the answer model is involved.2
Calibrated so on-topic questions score above 0.70 and off-topic ones fall to 0.48–0.59.3
Gemini and OpenAI, switchable for both embeddings and answers.
DOCX, PDF, TXT and Markdown, up to 20 MB each.
Keep exploring.
- 1.Document text is sent to the AI provider you choose, Gemini or OpenAI, to create embeddings. When a question is asked, only the passages that pass the relevance gate are sent along with it to write the answer.
- 2.When no passage clears the relevance gate, a fixed “not found” reply is returned and the language model is not called.
- 3.Calibrated for the default embedding model, gemini-embedding-2, so that on-topic questions score above 0.70 and off-topic ones between 0.48 and 0.59. Scores depend on the model, so the cut-off should be re-checked after switching.