Discovery products need a knowledge contract
Content and knowledge products help people find, understand, compare and act on material they did not create in the current moment.
- Search and recommendation
- AI interaction
- Content operations
- Software delivery
The corpus changes the product
Different products organise different knowledge contracts. The team must decide what counts as a document, passage, version, collection, topic, source and recommendation before it can judge whether search is good.
- 01
Publishing and media products
They balance timeliness, editorial structure, identity and audience journeys.
- 02
Research and scientific platforms
They require precise metadata, versions, methods and citation paths.
- 03
Organisational knowledge products
They inherit permissions, ownership and changing source systems.
- 04
Learning libraries
They connect discovery to sequence, practice and progress.
- 05
Developer documentation
It connects concepts, examples, versions and executable work.
- 06
AI retrieval products
They create a new synthesis layer whose source coverage must remain reviewable.
Is the corpus ready for search or AI-assisted answers?
Readiness depends on ownership, structure, freshness, access rules, duplication and whether important claims can be traced to a source. Gaps do not always block a bounded pilot, but they limit what the product may promise.
The same readiness decision names the authoritative source when versions conflict and the review path when no source can settle the answer.
From corpus to action without losing provenance
The path begins with intent and source eligibility, then moves through retrieval, ranking, presentation, interpretation and action. Search can rank material; AI can assemble an answer.
In both cases, the reader needs claim-level support, permission and version boundaries, visible uncertainty and a correction route that updates indexes and dependent answers.
Search, browse and recommendation serve different intent
Search helps when a person can express what they need. Browse helps when they need to understand the available space. Recommendation helps when the product uses context and an objective to bring something forward.
A useful experience can combine these modes without presenting every route as one search box. It should preserve:
- what the system searched;
- how filters changed eligibility;
- why a result or recommendation appeared;
- which source and version the result represents;
- what the person can preview before committing;
- how feedback changes later discovery.
In the Eden AI case, Tcules brought template discovery, AI assistance and from-scratch workflow creation into one entry model while preserving their distinct modes. Preview and documentation access kept technical context near selection.
AI reduces reading cost and adds a coverage problem
Summaries, answer generation and conversational retrieval can help someone move through a large corpus. They can also compress disagreement, omit relevant sources or present a partial answer with the tone of a complete one.
The article on generative AI product patterns distinguishes retrieval and synthesis from broader delegated action.
The product should expose source coverage, permission and version where they affect the decision. For high-consequence use, citation must connect the claim to the passage rather than only provide a list of documents. Correction and feedback should not rewrite the underlying source.
The W3C PROV-O recommendation (opens in a new tab) provides one formal vocabulary for representing provenance. Tcules uses standards according to the product context and does not imply that every content platform needs that implementation.
How the work comes together
Tcules can combine corpus and object modelling with information architecture, metadata, taxonomy, search and recommendation UX, and AI evidence and control. Delivery can extend to responsive reading, design systems, frontend, backend, APIs and integrations.
The public proof on this page is limited to Eden AI's workflow discovery and The Wandering Raven's learning-content and Moodle implementation context.
Bring the corpus, the person's decision and one result that is currently difficult to trust or act on. That is enough to locate the discovery problem.
Bring the corpus, the person's decision and one result
That is enough to locate the discovery problem.