query-knowledge

name	query-knowledge
description	Query existing LightRAG knowledge in a claw-scoped workspace without ingesting or mutating documents.
homepage	https://github.com/trinityagi/trinity
metadata	{"openclaw":{"emoji":"🔎"}}

Query Knowledge

Search, compare, inspect, and summarize knowledge that is already indexed in the active openclaw LightRAG workspace.

Purpose

Use this skill when a claw needs retrieval-backed answers from existing knowledge without performing ingestion, re-indexing, or document mutation.

Examples:

• search an indexed handbook for applicable requirements
• compare two indexed documents against a question
• inspect document status, chunks, or graph labels before analysis
• retrieve evidence for a downstream compliance or reasoning skill

Responsibilities

• list existing indexed documents
• inspect document status and chunks
• run scoped retrieval queries
• run scoped compare queries
• inspect graph labels or graph slices when graph knowledge is useful
• return evidence-backed retrieval results for downstream workflows

Non-Responsibilities

• do not upload, register, ingest, or delete documents
• do not create or alter workspaces
• do not invent knowledge that is absent from indexed sources

Workspace Contract

Knowledge queries must stay in the target claw's derived default workspace, not an arbitrary caller-defined workspace.

Default convention:

• tenant_<tenantId>__claw_<openclawId>

Rules:

• derive the active workspace from the current openclaw context
• do not switch to alternate workspaces unless the platform explicitly does so
• cross-workspace query and compare are not allowed

Query Contract (same workspace only)

Read/query through auth-service scoped endpoints for the active OpenClaw:

• GET /auth/openclaws/:id/lightrag-documents
• GET /auth/openclaws/:id/lightrag-documents/:documentId/status
• GET /auth/openclaws/:id/lightrag-documents/:documentId/chunks
• GET /auth/openclaws/:id/lightrag-scope
• GET /auth/openclaws/:id/lightrag-search
• GET /auth/openclaws/:id/lightrag-label-search
• GET /auth/openclaws/:id/lightrag-graph
• POST /auth/openclaws/:id/lightrag-query
• POST /auth/openclaws/:id/lightrag-compare

Auth Modes

Use one of these auth modes depending on caller context:

• user-scoped auth via valid JWT when running from the app UI or user session
• delegation token via X-Trinity-Delegation when the caller already minted a scoped delegation token
• claw-runtime service auth via X-OpenClaw-Gateway-Token when running inside the claw for the same openclaw

Operational rules:

• for claw-runtime calls, prefer OPENCLAW_ID env (UUID) for :id; name aliases may resolve but UUID is canonical
• use TRINITY_AUTH_SERVICE_URL as the base URL for /auth/openclaws/* routes
• do not send /auth/openclaws/* routes to the OpenClaw gateway port
• workspace is enforced server-side; do not try to override it in request payloads

Relationship to Ingest Knowledge

• ingest-knowledge prepares and indexes knowledge
• query-knowledge only reads and compares knowledge that is already present
• downstream skills can depend on query-knowledge without taking on ingestion responsibilities

Relationship to OpenClaw Memory

• this skill uses the LightRAG service and its own workspace-scoped knowledge store
• OpenClaw memory plugin settings do not configure these retrieval endpoints
• both systems are complementary: memory plugin supports agent memory, LightRAG supports indexed document retrieval

Expected Input

Provide at minimum:

• one active openclaw workspace containing indexed knowledge
• one query, comparison task, document id, or label to inspect

Optional:

• preferred handbook version or document ids when multiple sources exist
• retrieval mode, top-k, or graph label when the caller needs tighter control

Output

Return:

• retrieved evidence and citations
• matched documents or document ids
• relevant chunks or references
• compare results when used
• graph labels or graph metadata when used
• concise summary grounded in indexed evidence

Notes

• do not ingest or re-index documents here; use ingest-knowledge separately when knowledge is missing
• do not claim missing knowledge is present if retrieval returns nothing
• prefer exact quotes and references over paraphrase when supporting a downstream decision