poly-gents Vectorize
Purpose
Retrieve relevant passages from vector-indexed content: phrase queries clearly, treat chunks as evidence, summarize with brief citations—using tool names from the Vectorize MCP server (list_tools).
When to Use
- Semantic search over a knowledge base or indexed corpus.
- “What do we know about …” questions needing retrieved context.
- Finding notes or docs by meaning, not exact filename.
When NOT to Use
- Authoritative Syncolab platform docs → syncolab-living-docs (SKS).
- Tenant long-term agent memory (CML REST/hydration) → syncolab-cognitive-memory.
- Fabricating answers without running a query tool.
Expected Outcome
- One focused natural-language query first; refine only if results are weak.
- Response summarizes chunks; does not dump raw text walls.
- Tool names and parameters from live MCP schema.
Inputs to Gather
- Clear natural-language query (avoid overly short keywords).
- Optional filters if the MCP schema exposes them (collection, tenant, etc.).
Workflow
- Discover Vectorize tools via MCP
list_tools; read schema before call. - Run a single well-phrased query.
- If results are off-topic, refine query once with more context.
- Synthesize answer with short citations to chunk sources/ids when present.
Domain rules
- Phrase queries clearly in natural language.
- Chunks are evidence—summarize and cite, do not paste entire indexes.
- One focused query first before iterating.
Main tools
- Vectorize MCP query/retrieve tools (names from
list_toolsat runtime).
Examples
Q4 launch: Query "Q4 product launch"; return concise summary with chunk references.
API auth flow: Query "API authentication flow"; present matching passages briefly.
Tool Availability Rules
| Access | Behavior |
|---|---|
| Full MCP access | Query and summarize results. |
| Read-only | Query only. |
| No Vectorize MCP | Do not invent retrieved text. |
Related tool sets
openai
Review / Decision / Execution Criteria
- Prefer precision over dumping all chunks.
- Distinguish low-confidence empty results from “no data.”
Output Format
- Query used.
- Summary with citations.
- Gaps or low relevance.
- SKS/CML skill if user needs platform canonical docs or tenant memory.
Quality Bar
- Grounded answers only in retrieved chunks.
- No hallucinated citations.
Safety and Boundaries
- Do not leak sensitive chunk content beyond user need.
- No secrets in queries.
Escalation / Dispatch Rules
- Platform truth → syncolab-living-docs.
- Agent memory layer → syncolab-cognitive-memory.
References
- Legacy:
skills/old_skills.json(vectorize). skills/skill.instruction.md,skills/meta.instructions.md