| name | openmark-fast-search |
| description | Quick one-shot lookup over Ahmad's OpenMark bookmark knowledge graph (15k+ items from Edge, LinkedIn, YouTube, Raindrop). Use whenever Ahmad asks "find my bookmarks on X", "what did I save about X", "anything in OpenMark about X", or any single-topic search where he wants results in under 10 seconds. NOT for newsletter drafting or deep research — for those use openmark-deep-research or openmark-newsletter. Triggers on "find my", "what did I save", "openmark", "my bookmarks", "search bookmarks". |
| metadata | {"type":"search"} |
OpenMark — Fast Search
Single-shot search over Ahmad's personal knowledge graph. Optimised for "I just want results, now."
The recipe (in this exact order)
- One call:
search_semantic(query=<user query>, n=12).
- Inspect the returned
hits[]. Each hit has url, title, similarity, category, source, tags, community_id.
- No second tool call unless you see one of these failure signals:
total_found == 0 → fall back to find_by_tag if the query looks like a single token, else search_by_category if you can map it to a canonical category.- First three hits all have
similarity < 0.55 → escalate to search_by_community (same query).
- User explicitly named a source ("youtube", "linkedin") → use
search_youtube or search_linkedin instead of search_semantic from the start.
Output format
Numbered list. Each line is: N. Title — URL — short why-relevant (one phrase, max 12 words).
The "why" comes from category, source, and tags — not invention. If a hit's title is bare/generic, lean on tags or source.
Do not add a leading "Here's what I found" or any preamble. Get to the list.
Do not invent URLs. Every URL you emit must appear in the hit you cite.
Do not truncate URLs. Full URL or nothing.
When to bail out
If search_semantic returns 0 hits AND the fallback returns 0 hits, say so plainly:
No matches in OpenMark for <query>. Try a broader term, or use openmark-deep-research for cross-source expansion.
Don't apologise. Don't recommend WebFetch — that's deep-research territory.
Example
User: "find my bookmarks on prompt caching"
search_semantic(query="prompt caching", n=12) → 12 hits.- Render:
1. Prompt Caching with Claude 3.5 — https://docs.anthropic.com/.../prompt-caching — official docs, AI Tools
2. Reducing Cost with Prompt Caching — https://blog.anthropic.com/... — engineering blog post
3. How Vercel uses prompt caching — https://vercel.com/blog/... — production case study
...
That's it. Done in one tool call. Move on.
