Ask People Team

An employee asking about parental leave, notice periods or benefits is not browsing. They are making a decision, often a consequential and personal one, and they will act on whatever the assistant says. A confidently wrong answer about entitlement is worse than no assistant at all, because it carries the organisation’s authority.

That reframes the engineering. The hard requirements are not fluency or coverage. They are: answer only from the company’s own documents, make it clear where an answer came from, and have a way of knowing whether retrieval quality is getting better or worse.

The assistant lives in Slack. An HR knowledge tool that requires finding a portal, logging in and searching is a tool people use once. The question gets asked in the place people are already complaining that they cannot find the answer.

HR documents are long, structured and full of clauses that only make sense in context. Split them badly and retrieval returns a fragment that reverses the meaning of the paragraph it came from.

The ingestion pipeline therefore treats chunking as a first-class decision rather than a default. More than one strategy is implemented — sentence-aware and recursive — so the approach can be chosen and compared rather than assumed.

Dense vector search finds passages that mean the same thing in different words, which is what employees actually type. Sparse keyword matching finds exact terms — a policy name, a form number, a specific benefit — where semantic similarity is not enough. Using both covers the failure mode of each.

The vector store runs as a deployed service with its own storage and coordination components rather than as a library, so the index is operable, inspectable and separable from the application that queries it.

Retrieval quality can be improved but not guaranteed, and HR answers carry consequences. A human review step sits before an answer reaches an employee, which changes what the model has to be: not the final authority, but a well-grounded first draft with its sources attached.

Retrieval systems degrade quietly. A chunking change, a new document set or a model swap can improve some questions and break others, and nobody notices until an employee is told something wrong.

A dedicated evaluation service runs the system against a curated golden dataset of questions with known good answers, and reports the result to a dashboard. That turns "does this change help" from an opinion into a measurement.

When someone reports a bad answer, the question is which stage failed: routing, retrieval, or the model’s use of what it retrieved. Tracing across the pipeline records each stage, so a specific answer can be reconstructed rather than guessed at.

The assistant, the ingestion pipeline, the evaluation service, its dashboard, the vector store and its supporting components, the relational database and the tracing service are all separate services, composed so the whole system runs locally as well as deployed.