How do I install the Lightrag Deep Research skill?

Run npx skillmds@latest add olafgeibig/lightrag-deep-research in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Lightrag Deep Research safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Lightrag Deep Research?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Lightrag Deep Research free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Lightrag Deep Research?

olafgeibig (@olafgeibig) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Lightrag Deep Research by olafgeibig

Name: Lightrag Deep Research
Author: olafgeibig

Lightrag Deep Research

Deep research workflow over an internal LightRAG knowledge base (no web search). Use when you need an in-depth investigation with a structured plan, iterative sub-questions, evidence gathering via LightRAG queries (local/global/hybrid/mix/naive), synthesis, and KB-citations. Best for: architecture/security research, incident retrospectives, design decisions, policy interpretation, and any question that should be answered primarily from your organization’s knowledge base.

olafgeibig 91e72de 3.3 KB Updated Sep 17, 2026

File contents

LightRAG Deep Research

Implement a “deep research” pattern, but replace web search with LightRAG KB search.

Goals:

Create a research plan (sub-questions, assumptions, scope).
Run multiple LightRAG queries to gather evidence.
Synthesize with KB citations so the user can verify.
Be explicit about gaps / missing KB coverage.

Inputs to clarify (ask if missing)

The research question (what decision or deliverable is needed?).
Desired depth (quick / standard / deep).
Constraints: time, doc type, timeframe, system/component scope.
Preferred KB server alias (if multiple).

Workflow (deterministic)

1) Restate + scope

Restate the question in one sentence.
State assumptions and what is in scope / out of scope.

2) Decompose into sub-questions

Produce 5–12 sub-questions depending on depth.
Tag each as: context, facts, decision criteria, risks, open items.

3) Evidence gathering (LightRAG)

Use the existing lightrag skill script to query LightRAG repeatedly.

Prefer hybrid or mix for broad discovery; use local for grounded retrieval; use global to explore wider connections.
For each sub-question:
- Run at least 1 query; for high-impact items run 2–3 variations with different phrasings.
- Capture the returned context and also keep a short citation stub (query + mode + timestamp + server alias).

Commands (use the existing LightRAG skill script):

Query and return response: python3 ~/.agents/skills/lightrag/scripts/query_lightrag.py query "<text>" --mode hybrid
Query and return only context: python3 ~/.agents/skills/lightrag/scripts/query_lightrag.py query "<text>" --mode hybrid --only-context

4) Synthesis

Gate: do not write the final answer until you have gathered enough evidence.

Default minimum: 5 distinct LightRAG queries (unless the question is trivial or the user asked for a quick answer).
Summarize key findings.
Highlight agreement vs. contradictions.
Identify what is supported by the KB vs. inferred.

5) Output

Use this structure:

## Executive Summary

## Research Plan (what was asked / how investigated)

## Key Findings
- Finding … [KB-1]

## Detailed Analysis
### Subtopic …

## Risks / Caveats

## Gaps & Next Steps

## KB Citations
[KB-1] server=<alias> mode=<mode> query="..." (context excerpt)

Notes:

KB citations are not “paper citations”; they are reproducible retrieval references.
Keep excerpts short; prefer linking to internal doc titles/ids when present in context.

When to stop

Stop evidence gathering when:

Additional queries repeat the same context,
or remaining open questions cannot be answered from the KB.

Bundled resources

None. This skill is intentionally composable: it instructs how to use the existing lightrag skill/scripts.

---name: lightrag-deep-researchdescription: Deep research workflow over an internal LightRAG knowledge base (no web search). Use when you need an in-depth investigation with a structured plan, iterative sub-questions, evidence gathering via LightRAG queries (local/global/hybrid/mix/naive), synthesis, and KB-citations. Best for: architecture/security research, incident retrospectives, design decisions, policy interpretation, and any question that should be answered primarily from your organization’s knowledge base.---# LightRAG Deep ResearchImplement a “deep research” pattern, but **replace web search** with **LightRAG KB search**.Goals:- Create a research plan (sub-questions, assumptions, scope).- Run **multiple** LightRAG queries to gather evidence.- Synthesize with **KB citations** so the user can verify.- Be explicit about gaps / missing KB coverage.## Inputs to clarify (ask if missing)- The **research question** (what decision or deliverable is needed?).- Desired **depth** (quick / standard / deep).- Constraints: time, doc type, timeframe, system/component scope.- Preferred KB server alias (if multiple).## Workflow (deterministic)### 1) Restate + scope- Restate the question in one sentence.- State assumptions and what is **in scope / out of scope**.### 2) Decompose into sub-questions- Produce 5–12 sub-questions depending on depth.- Tag each as: *context*, *facts*, *decision criteria*, *risks*, *open items*.### 3) Evidence gathering (LightRAG)Use the existing `lightrag` skill script to query LightRAG repeatedly.- Prefer `hybrid` or `mix` for broad discovery; use `local` for grounded retrieval; use `global` to explore wider connections.- For each sub-question: - Run at least 1 query; for high-impact items run 2–3 variations with different phrasings. - Capture the returned **context** and also keep a short **citation stub** (query + mode + timestamp + server alias).Commands (use the existing LightRAG skill script):- Query and return response: `python3 ~/.agents/skills/lightrag/scripts/query_lightrag.py query "<text>" --mode hybrid`- Query and return only context: `python3 ~/.agents/skills/lightrag/scripts/query_lightrag.py query "<text>" --mode hybrid --only-context`### 4) SynthesisGate: do **not** write the final answer until you have gathered enough evidence.- Default minimum: **5 distinct LightRAG queries** (unless the question is trivial or the user asked for a quick answer).- Summarize key findings.- Highlight agreement vs. contradictions.- Identify what is *supported by the KB* vs. *inferred*.### 5) OutputUse this structure:```markdown## Executive Summary## Research Plan (what was asked / how investigated)## Key Findings- Finding … [KB-1]## Detailed Analysis### Subtopic …## Risks / Caveats## Gaps & Next Steps## KB Citations[KB-1] server=<alias> mode=<mode> query="..." (context excerpt)```Notes:- KB citations are not “paper citations”; they are reproducible retrieval references.- Keep excerpts short; prefer linking to internal doc titles/ids when present in context.## When to stopStop evidence gathering when:- Additional queries repeat the same context,- or remaining open questions cannot be answered from the KB.## Bundled resourcesNone. This skill is intentionally composable: it instructs how to use the existing `lightrag` skill/scripts.

Frequently asked questions