Kora — Internal AI Knowledge Assistant

Year

2026

Client

Meridian Works (anonymised, per the published case study)

Role

Founder & Lead Engineer, Refactrd — designed and built the full system

Built at

Refactrd

Read Full Case Study

Overview

Kora is an internal AI knowledge assistant built at Refactrd for a growing professional services firm. Employees were losing time searching for information that already existed, and new hires took too long to get up to speed. Kora answers questions using retrieval-augmented generation over the organisation's indexed documents rather than a model's general training data, and every answer includes a citation so the employee can verify the source. If Kora does not have enough information to answer confidently, it says so and points the employee to the right person or team instead of guessing.

Challenges

New hires struggled to get up to speed on policy, managers answered the same questions repeatedly, and the knowledge existed in documents nobody could find in time. The demo in the public case study runs on a representative, anonymised set of documents to protect client confidentiality; the engineering, infrastructure and AI are identical to what is live for the real organisation.

Solutions

Document ingestion pipeline: uploaded documents are chunked, embedded with OpenAI's text-embedding-3-small (1536-dim), and stored in Supabase using the pgvector extensionQuery pipeline: validate the query, check a semantic cache (0.92 cosine similarity threshold) before any model call, run a pgvector similarity search, then score confidence and route low-confidence queries to an 'uncertain' response instead of a guessGrounded generation: retrieved chunks and the question are passed to Claude Sonnet 4 with a system prompt instructing it to answer only from the provided context and cite its sources, streamed to the employee via Server-Sent EventsAn admin suite for document management plus analytics: total queries, a 14-day rolling query chart, a confidence breakdown, and an 'uncertain query' log that surfaces exactly which topics the knowledge base cannot yet answer

Results

After indexing the organisation's documents and running for several weeks, the uncertain-query log surfaced real documentation gaps (expense reimbursement timelines, equipment requests, complaint escalation) that did not exist as written policy before and have since been written. Typical build timeline for a system like this is two to four weeks from document ingestion to a live, usable assistant.