From hours of waiting
to answers in seconds.
Grounded answers layer. A layer on the company’s existing AI inference platform that kept technical documentation and source code synced as dedicated knowledge bases, and answered from them with the source attached. Field teams got real customer answers in Slack in seconds instead of hours.
- Context
- Built by Francois as part of his work at a large enterprise software company, on its existing AI inference platform.
- Service
- Knowledge sync and ingestion (documentation, source code), retrieval and citation grounding, evaluation harness, data governance, Slack delivery.
- Industry
- Enterprise software
- Specialty
- Build and integrate AI
- Status
- Live. More than twenty knowledge bases, open to the whole company.
Overview
Inside a large enterprise software company, the answers people needed already existed in documentation, source code, tickets and past responses. Finding them was the work. The company already had an AI inference platform. What it lacked was the layer that turned scattered technical sources into dedicated knowledge bases and kept them in sync. That layer was built on top, through the platform’s APIs, so nothing that already worked was rebuilt. Several delivery mechanisms shipped; the one that won on usage was Slack channels, where field teams answered real customer questions, fast.
Challenge
Technical questions typically took 2–3 hours to receive a response, based on communication histories before the workflow. The time was in the finding, the asking and the waiting, and the same question was answered many times, differently. An answer without a source could not be trusted with a customer. A knowledge base that went stale the week after it was loaded was worse than none. A second inference stack would also have added infrastructure and operating overhead.
Solution
A shared ingestion layer any internal team could plug a knowledge base into, with automated syncs from documentation and source code to keep the knowledge bases current: content transformers and a shared source factory to normalise what came in, chunking tuned for retrieval quality, link resolution and canonical URL normalisation so every passage traced back to one address, and provenance stamped on every chunk.
On top, retrieval with citation grounding, safety through system prompt engineering, and an evaluation harness using an LLM as judge to measure retrieval quality and catch regressions. Automated audit and data governance decided what each team’s content could be used for, and by whom. The inference itself stayed on the company’s platform.
Results
From hours of waiting to answers in seconds.
- Response time before
- Typically 2–3 hours
- AI response time after
- Typically 5–10 seconds
- Adoption
- About 600 users and 2,000 questions over five months
- User satisfaction
- Above 80% among rated responses
- Feedback coverage
- Approximately 40% of questions received a user rating, roughly 800 responses
Response times reflect before-and-after email, Slack and other communication histories. Satisfaction reflects submitted user ratings; unrated responses are excluded. Time to full resolution was not consistently measured. Adoption figures cover one knowledge base, and feedback coverage is approximate.
- Presales, post-sales and go-to-market teams received sourced technical answers in Slack while working with customers.
- Built on the existing inference platform, with no second inference stack to run.
- Retrieval evaluation, provenance and audit capabilities supported ongoing quality checks. User satisfaction ratings were separate from automated evaluation.
Relevant experience for knowledge retrieval, system integrations and AI evaluation.