rag-onboard-context

rag-onboard-context — know your corpus before you use it

When to use this skill

Run this skill:

• At the start of any new conversation where knowledge-rag is available and the user is about to ask substantive questions
• When the topic shifts significantly (from security to infrastructure, from dev to research)
• After a major reindex (the corpus content changed under you)
• When the user says "I just indexed new docs" or similar

Do NOT run repeatedly — once per session is usually enough. The query cache applies to searches; it does not cache every inspection tool.

---

What this skill commits to

Before diving into task-specific work, the agent gathers a mental map of the corpus:

• How big — chunk count, document count, cache health
• What is in it — which categories exist, roughly how many docs per category
• What flavor — probe 1–2 topics the user is likely to ask about, see if there are hits

That mental map informs every subsequent decision: whether to search first, whether to fall back to the web, how to phrase queries.

---

Steps

1. Get index health:

   get_index_stats()
   

   Read the nested stats object: total_documents, total_chunks, query_cache.hit_rate, embedding_model, and embedding_dim.

2. Enumerate categories:

   list_categories()
   

   Note: which categories exist and how many docs each has. This tells you the corpus's shape.

3. Probe 1–2 topics the user is likely to touch. If the user's first message mentions a domain, probe it. Otherwise, probe the top 2 largest categories:

   search_knowledge(query="<category-typical-term>", max_results=3, snippet_mode=true)
   

4. Optionally, if you need concrete file names, call:

   list_documents(category="<top-category>")
   

   This tool has no limit argument and returns the whole category. For a large category, prefer a targeted search_knowledge call with max_results=3.

5. Store the summary internally — do not necessarily surface it to the user unless they ask. The value is that YOU now know:

   • "The corpus has 15K chunks across 8 categories"
   • "Biggest categories are security (5K), development (3K), research (2K)"
   • "Search for mitre attack returned 3 solid hits — MITRE content is well indexed"
   • "Search for kubernetes returned nothing — try synonyms or inspect relevant categories before concluding the topic is absent"