Services

AI Search and Recommendation

Design search and recommendation around corpus, access, provenance, relevance, uncertainty, feedback and the decision that follows.

  • AI product UX
  • Search design
  • Recommendation systems

Help people move from a question to inspectable evidence

AI search can retrieve passages, generate a synthesis or recommend what to do next. Those are different product responsibilities. A fluent answer is not the same object as a source, and a high relevance rank is not an explanation of truth.

Tcules designs search and recommendation around corpus, access, provenance, relevance, uncertainty, feedback and the decision that follows.

Start by defining what the person is trying to find

The same query can represent several needs:

  • Find a known item.
  • Understand a topic.
  • Compare options.
  • Retrieve evidence for a decision.
  • Resume previous work.
  • Discover a useful capability.
  • Receive a recommendation under a stated objective.

The product may combine keyword, semantic, filtered and conversational interaction. It should not hide corpus and filter boundaries simply because the query is expressed in language.

From query to supported action

  1. 01

    User need or objective

    The experience begins by identifying what the person is trying to find or decide.

  2. 02

    Permitted corpus

    Permission and filters determine which corpus items are eligible before retrieval or synthesis begins.

  3. 03

    Retrieved passage

    Ranking selects passages, not truth. Retrieved passages retain source and version so evidence can be inspected.

  4. 04

    Synthesis

    Each consequential claim should connect to supporting passages and distinguish missing evidence from negative evidence.

  5. 05

    Recommendation and alternatives

    A recommendation names the objective, basis and alternatives before action.

  6. 06

    Feedback

    Feedback can improve discovery without rewriting source history.

Recommendation needs an objective and a correction path

A recommended template, account, article, product or action is optimising something. The interface should state enough of that objective for the person to judge whether the result is useful. Relevant context may include role, prior activity, constraints, recency or product state.

People should be able to correct material context, dismiss an unsuitable result and understand whether feedback affects only the current decision or future recommendations. High-consequence recommendation may require explicit evidence and human review rather than a generic “Because you viewed” explanation.

Evaluation is layered

  1. 01

    Corpus and access

    Was the right material eligible and unauthorised material excluded?

  2. 02

    Retrieval

    Were the relevant passages found?

  3. 03

    Support

    Does the generated claim follow from those passages?

  4. 04

    Synthesis

    Is the answer complete, bounded and useful?

  5. 05

    Recommendation

    Does it serve the stated objective and expose alternatives?

  6. 06

    Decision

    Can the person act, correct or recover appropriately?

A fluent response can fail because the right evidence was never retrieved. A correct retrieval can still be synthesised badly.

Proof from a developer product

The Eden AI case shows Tcules unifying template search, AI assistance and from-scratch creation while preserving their modes. Preview and documentation access kept technical evidence near selection, and the chosen entry continued into one workflow-builder context.

The case supports search and workflow interaction. It does not establish backend retrieval, model quality, provider coverage or conversion outcomes.

Tcules can continue from product modelling and interaction into coded prototypes, retrieval and model integration, evaluation and software delivery according to scope.

Design search around evidence and action

Bring the search task, corpus boundary and decision context so the service can start from the right product responsibility.