Context

Organizational knowledge is spread across documents whose wording, structure, and relevance vary. Staff need a useful answer and a way to inspect the material behind it. Gilles delivered production retrieval and source-linked answers over rich-format organizational documents. A separate staff knowledge assistant reached pre-production, with rollout underway when the engagement ended. These are different delivery states; the existence of a working retrieval backend does not establish a completed production rollout.

Approach

Ivory index cards in a green leather tray

The work included document ingestion, extraction, metadata-filtered retrieval, and conversational answers tied to sources. A separate knowledge-base ingestion effort supplied product, service, and departmental material for staff assistance. Common design choices involve preserving useful document context, selecting relevant passages, and keeping references attached to the response. Predictable data ingestion explains the upstream processing concerns. Document structure matters here: a relevant passage may depend on its heading, an adjacent qualification, or the context of a table rather than an isolated fragment.

Quality controls

Verified quality work included retrieval evaluation, experiment tracking, annotation-based validation, and user ratings with optional comments across the relevant engagements. Feedback informed prompt refinement. More generally, answer quality should be separated from retrieval quality, and absent evidence should be visible to the user. The article on adaptive retrieval and measurable answers discusses when additional search steps are worth evaluating. These recommendations do not imply that every example used the same evaluation process.

Outcomes

Paper folder and wax seal beside a folded letter

Production delivery established source-linked organizational retrieval. The separate staff assistant established a working pre-production capability. Document-grounded compliance analysis shows another use of retrieved evidence, with a different interpretation task and review boundary.

  • Delivered production document retrieval and source-linked answers.
  • Developed a separate pre-production assistant with user feedback collection.