# Knowledge and procedural Skills Read this when configuring QA/factual material or separating it from SOP. Protocol source: https://api.latenth.com/docs/admin.openapi.json and https://api.latenth.com/docs/realtime-events.html. - Prompt: identity, scope, style and common constraints. - Skill: business procedure, decision branches, required fields and stop conditions. `preload:true` loads the text; `false` lets the model load it as needed. - Knowledge: versioned Markdown with optional text `resources`. No mandatory QA schema, embeddings or prebuilt vector index. Clear headings, conditions, questions and answers help runtime original-text search. Preserve source facts. ## Configure 1. `PUT /voice/v1/knowledge/{id}/versions/{version}` with `markdown` and optional `resources`. Include name/description Markdown frontmatter. Versions are immutable; use a new version when content changes. `GET` the same path returns full text and resources, not just a description. 2. For a saved custom model, read the latest draft/revision, preserve all unrelated fields and set `config.knowledge_refs:[{id,version}]`. Keep SOP in `config.skills` as saved `{id,version,preload}` references created with `PUT /voice/v1/skills/{id}/versions/{version}`; custom models reject inline Skill markdown at validate. Validate and publish only within the requested scope; use the actual returned model ID. Existing published versions are unchanged. 3. For a temporary predefined-model session, use `session.latentx.knowledge_refs:[{id,version}]`. Omitted references preserve existing selection; `[]` clears them. All references are tenant scoped. 4. `search_knowledge` is mounted and executed by the platform. The model may provide up to 3 independent `queries`, each with a required `query`, optional attached `knowledge_ids` and `limit` of 1–3. The legacy single `query` form remains supported; do not mix it with `queries`. Each call supplies a brief context-appropriate `waiting_message`, separate from the customer-facing answer; providing it does not guarantee playback. Results contain source-grounded `matches` and per-query references. Customers do not execute this tool or register its schema themselves; no separate public search HTTP endpoint is provided. 5. Do not configure an external `find_qa` or duplicate `search_knowledge` to use this built-in capability. `tool_choice:none` disables search; it is not a performance optimization for factual answers. ## Verify Confirm the selected knowledge versions, a question with evidence, a question absent from the corpus, and isolation from unattached knowledge. Record status/response IDs and latency, not customer source text in public reports. A model claiming it searched is not proof; inspect available tool events or platform evidence. Hosted search may run internally without customer Function events.