Self-learning

Open a datapack and choose the Self-learning tab. It shows whether the searches agents ran found enough context in this datapack. The tab is off unless the deployment turns it on, and it is available only for the datapack owner and a maintainer.

The agent can still reply when a search falls short. On a partial match it uses what the datapack returned and says what is missing. That missing part was never in the datapack. The agent may already know it, or it may look it up with another tool, such as a web search. The score counts only what came back from the datapack. High means that result was enough to answer from. Below that, the reply can still look complete.

Store a fact once. Questions that need that fact can find it without a separate sentence for every wording. The Self-learning tab shows which searches come back short again and again. Add the missing fact, then check that the same search does better.

Watch your agent learn is a fact an agent stores and a later session reads back. Learnings moves a private memory into the knowledge the team shares. Self-learning reads searches that have already run and shows whether what came back from the datapack was enough. It opens on Summary. Observe, Analyze, and Improve stay empty until an agent has called context_search on this datapack.

Observe

Observe answers what was searched, and how it went.

Each row is one context_search: the query, when it ran, and a quality band from the best result.

The list defaults to the last 14 days. The search log is kept for 30 days. A row appears once an agent calls context_search on this datapack. A search run to check a fix, from Improve, is not listed here.

Analyze

Analyze answers why searches fell short. High searches have no reason. For the rest, the first matching reason wins, in this order.

Two more reasons describe the datapack as it is now, not the search as it ran: Datapack is empty (no memories or documents yet) and Indexing in progress (content was added and is not searchable yet).

Improve

Improve answers what to add, and whether it helped.

Improve does not add a memory or a file. Add the missing fact from Memories, or the missing document with Add Knowledge, then come back here.

Pick a search that fell short and click Run the prompt again. The first run saves that search as a test prompt. A datapack holds up to 50 saved prompts. The run repeats the search and charts the best score against the earlier run. Each run counts as one search. If the new content is still indexing, the page says so: run the prompt again shortly.

Summary also lists what would raise the share of high matches: add the first memory or document, connect an agent so searches exist, fill a coverage gap, clarify a vague prompt, refresh an out-of-date source, or check a source that failed.

Check a fix from an agent

An agent uses four tools, which a deployment registers only when self-learning is on. The datapack's owner or a maintainer can call them. selflearning_verify needs a key that can write. The other three need a key that can read.

1. selflearning_status reads the quality and the open to-dos. Start here. selflearning_status and selflearning_shortfalls default to 14 days and selflearning_query_log to 7, so pass the same last to all three.
2. selflearning_shortfalls lists the searches behind one to-do. selflearning_query_log reads the search log, newest first.
3. Fix the shortfall with memory_add, memory_update, or knowledgebase_upload_document, using a source you can name.
4. selflearning_verify replays that search and reports the match before and after. Run it after something has changed. Passing log_id saves the search as a test prompt. The result includes outcome (improved, no_change, worse, or no_baseline), indexing, and nothing_added_since_last_run. When nothing_added_since_last_run is true, the replay is not new. When indexing is true, run it again once that is false. You can replay about 10 times a minute on a datapack. Each person has their own allowance.

See Self-learning tools.