Avatar PKM Retrieval

Problem

When a user asks for notes, uploaded files, or avatar-specific reference material from an avatar’s personal knowledge base, a general-purpose answer is usually not enough. The content lives in a customer-managed PKM dataset, so the retrieval path must target a concrete dataset_id instead of relying on public knowledge or broad product context. This skill turns that workflow into a narrow, predictable flow: validate the input, call the PKM retrieval endpoint, and compress noisy results into a usable answer.

How it works

It is intended for cases where the user has already supplied a dataset_id or is willing to provide one. The main steps are:
- Load skills/pkm-retrieval/config.json to read base_url and the API credentials;
- Build the retrieval request from the user’s query and call POST /v1/datasets/{dataset_id}/retrieve;
- Treat UI Vector Search as search_method: "semantic_search" in this environment;
- Prefer the highest-scoring chunk, discard noisy PDF or image-preview fragments, summarize the relevant content, and mention source document names when useful.

Boundaries

Do not use it for general web or product-knowledge questions, or when only an avatar_id is available and no dataset_id has been provided. If the dataset_id is missing, ask for it instead of deriving an internal dataset from the avatar_id. Credentials must come from the config file and should never be hardcoded in skill instructions or request payloads.