Work
A grounded knowledge copilot for field service
A working prototype: retrieval over a curated technical corpus with cite-or-abstain grounding, a confidence score per answer, a resolve-or-escalate recommendation and human approval before any knowledge update is published.
- What it proves
- Grounded retrieval for settings where a confidently wrong answer is expensive.
- Method stages
- Built from
- Retrieval over a curated technical corpus
- Strict cite-or-abstain grounding
- Confidence scoring per answer
- Resolve-versus-escalate recommendation
- Agentic knowledge-base drafting with human approval before publishing
- Sectors
The problem
A technician facing an unfamiliar fault needs an answer that is right or an honest “I don’t know”. A plausible wrong answer can cost more than the downtime it was meant to prevent. General-purpose chat is built to sound helpful, which is the wrong default here.
What we built
The system retrieves from a curated corpus of technical manuals, maintenance history and resolved cases. Every answer must cite what was retrieved or abstain. It never answers from general knowledge and attaches a citation afterwards.
Each answer carries a confidence score and a recommendation: resolve it now, or escalate to a specialist.
Knowledge bases drift as equipment and procedures change, so the system drafts its own updates. When it finds a gap, or a resolved case worth keeping, it writes an entry in the corpus’s format and sends it to a person for approval. The system never publishes on its own, so every entry in the knowledge base has a human sign-off.
What it proves for a client
Retrieval can be designed for the case where a wrong answer is expensive. It takes abstention, a confidence score tied to an action, and a human gate on what enters the corpus.
Method stages
It sits at Stage 02 · Prove the value, where the evaluation, built from technician questions, is written before the system. Scaling to more equipment families is a separate decision in Stage 05 · Scale what works. See the KARPINITY Method.