cadau
Log inTry Cadau
ProductAppsSolutionsPlansDocsBlogForumAbout
Log inTry Cadau
← All docs

Indexed docs and feedback

Unify how-to articles and skill documents into a two-layer index, so the system can retrieve the right content quickly before answering or acting, then keep improving answer qualit

Source docs/en/site/mech-index-loop.md

Mechanism goal

Unify how-to articles and skill documents into a two-layer index, so the system can retrieve the right content quickly before answering or acting, then keep improving answer quality from real conversation feedback — a knowledge loop that can evolve over time.

Canonical-body rule: Human-written how-to / skill bodies are the authoritative original. The feedback loop does not rewrite those originals. Improvements land on overlay layers such as answer supplements, workspace skills, and retrieval indexes (see 记忆锻造/记忆锻造.md).

For implementation and runtime detail, see 记忆锻造/记忆锻造.md.

How documents are organized (two layers)

  1. First layer: root index

A catalog at the knowledge root that marks each theme subdirectory.

  1. Second layer: subdirectory index

Each theme folder has a second-level index that points to specific how-to or skill documents.

  1. Content files

Each theme may contain one or more documents with the real rules, steps, constraints, and examples.

  1. Answer supplements (overlay, optional)

Documents created after Memory Forge adopts a proposal, stored in a separate directory from the original. Retrieval merges them with the original; on conflict the original wins.

Retrieve and execute

  1. The user asks a question or requests an action.
  2. The system first searches the root index and picks candidate theme folders.
  3. It then searches those subdirectory indexes and locates one or more most relevant documents (including overlay supplements, if configured).
  4. Hit content is provided to the model as context.
  5. The model uses that context for two kinds of work:

- answering the user’s question; - executing the request by the rules (for example skill-driven actions).

User feedback and data

  1. Users rate whether the answer or the action was good.
  2. Users may offer a better method or extra explanation.
  3. The full conversation is recorded and archived.

Closed-loop improvement rules

  1. Conversation feedback should be analyzed on a schedule to find high-value improvements (MemoryForge scan).
  2. Suggestions should become new content or index improvements, not a patch of the original:

- missing content → create an answer supplement; - missing steps → create a workspace skill; - retrieval miss → improve index.json aliases / tags; - wording already stable → archive a verified answer.

  1. Changes to the original still go through human PRs / admin process.
  2. After updating overlays or indexes, keep watching whether answers and execution actually get better.

Closed-loop result

Through “index retrieval → answer/execute with context → user feedback → overlay/index update”, the system gets better with use. That is a durable, document-driven improvement loop, while human originals stay auditable and authoritative.

Related mechanisms