A recruitment email system whose retrieval actually retrieves

The problem: a model that answers questions it has no source for

Recruitment email handling meant answering the same candidate and process questions over and over, by hand. Automating that with a bare language model trades a time problem for a trust problem: a chatbot bolted onto an inbox will answer everything, including the questions it has no business answering.

The approach

A retrieval-grounded email system rather than a chatbot. Every design decision exists to keep answers traceable to a source:

1. Hybrid search, then rerank

   Answers are pulled from a curated FAQ knowledge base in a Supabase vector store, retrieved through a hybrid function that combines semantic and keyword matching, then reranked so the strongest source wins rather than the first plausible one.

2. Chunked per FAQ, not per token window

   A retrieved passage is a complete answer, not a fragment torn mid-thought.

3. Complete or flagged, never partial

   Multi-question emails are parsed question by question. If the system can answer all of them from sources, it replies. If it cannot, the email escalates to a person. A partial answer never goes out dressed as a full one.

4. A record manager keeps the index honest

   Indexing tracks what has been ingested, so re-indexing updates and deletes cleanly instead of silently accumulating stale duplicates. This is the part most RAG builds skip and then quietly rot on.

How it fails safely

Failure	What happens instead
Question with no source	Flagged to a person. The model never invents an answer to cover the gap.
Email with five questions, four answerable	The whole email escalates. Complete or flagged, never partial.
FAQ content changes	Re-indexing updates and deletes through the record manager. No stale duplicate ever outranks the current answer.
Wrong answer discovered	It traces to a specific FAQ record, so it is a fixable content problem rather than a mystery inside a model.

Result

• Repetitive recruitment email answering automated without giving up source grounding.
• Answers trace back to a specific FAQ record, auditable end to end.
• Indexing is re-runnable and self-cleaning, which is what keeps a RAG system trustworthy six months in.

Built by Hisham Hanawi. This is an AI in the workflow.

Related, but a different build: AutoRecruit moves candidates through a hiring pipeline. This system answers their questions from a grounded knowledge base.