PKM Knowledge Base Retrieval

Problem

When a user asks for material from an avatar's personal knowledge base (PKM) rather than public web or product documentation, ordinary search can mix customer-managed records with public information, PDF previews, and image fragments. test2333 narrows the task to a dataset-scoped retrieval flow: it treats dataset_id as the required entry point, calls the PKM retrieval API, and extracts the most relevant chunks instead of dumping raw noise to the user.

How It Works

The skill does not hard-code credentials or derive a customer-facing dataset_id on its own. It reads skills/pkm-retrieval/config.json for base_url, dataset_id, and api_key, then calls POST /v1/datasets/{dataset_id}/retrieve. The key steps are:
- Confirm the user has a dataset_id and a retrieval query;
- Treat UI Vector Search as the API's semantic_search case;
- Prefer the highest-scoring chunk and ignore noisy long-tail fragments;
- Return a concise summary, with source document names when useful.

If only avatar_id is available, the skill pauses and asks for the dataset_id rather than defaulting to the internal avatar_id -> pkb_{avatar_id} -> dataset lookup path.

Boundaries

It fits customer-maintained PKM datasets where the task depends on private reference material. It is not suitable for general knowledge questions, cases where no dataset_id can be provided, or workflows that should go through Avatar chat instead of direct dataset retrieval. The response style is concise: answer first, source second, and clearly say when nothing relevant is found rather than infer unsupported content.