PKM Knowledge Base Retrieval

Problem

When the target information is stored in an avatar PKM, not the web or product docs, raw retrieval can mix relevant chunks with noisy PDF or image-preview fragments. This skill fixes the workflow: ask for the missing dataset_id, call the retrieval endpoint, and summarize the best result without assuming internal mappings or hardcoding secrets.

How It Works

• Config: Load base_url and api_key from skills/pkm-retrieval/config.json; do not embed secrets in instructions.
• Input check: Require dataset_id and a retrieval query; do not derive it from avatar_id unless the user explicitly asks for the internal admin flow.
• Call: Use POST /v1/datasets/{dataset_id}/retrieve; UI Vector Search maps to semantic_search.
• Result handling: Prefer the highest-score chunk, ignore noisy PDF or image-preview fragments, and return a concise summary with source document names when helpful.

Boundaries

Best for customer-maintained PKM datasets. Not for general web or product questions, cases with only avatar_id, or tasks that belong in the Avatar chat workflow. If nothing relevant is found, say so clearly instead of guessing.