kg-reason

Files (1)

• SKILL.md

SKILL.md (first lines)

# KG Reasoning

Query the knowledge graph with natural language questions. Two modes: `ask` for quick KGQA
(retrieve → answer), and `reason` for deep LLM agent reasoning with tool-calling (iterative
graph exploration, section reading, hypothesis validation).

## Prerequisites

The knowledge graph must be built (`kg-build` skill). Best results with embeddings and closure:

```bash
drbrain index build && drbrain graph closure --mode hybrid
```

## Operations

### ask — Quick KGQA

Natural language question against the knowledge graph. Retrieves top-k relevant concepts and
edges, then generates an answer grounded in the graph:

```bash
drbrain ask "Is attention better than CNN for NLP?"
drbrain ask "what are the main approaches to knowledge distillation" --top 10
drbrain ask "how does BERT handle long sequences" --json
```

Options: `--top` / `-k` (concepts to retrieve, default 5), `--json`.

The answer includes provenance — concepts and edges used as evidence.

### reason — LLM Agent Reasoning

LLM agent with tool access to the knowledge graph. The agent can query concepts, traverse edges,
read paper sections (via PageIndex tree), and iteratively refine its answer:

```bash
drbrain reason "compare the efficiency of sparse vs dense attention mechanisms"
drbrain reason "what open problems remain in graph neural network explainability?"
```

**Bidirectional mode** (`-b`): The agent proposes hypotheses, validates them against graph
constraints (TBox/RBox), and revises. Useful for questions where contradictions or constraints
matter:

```bash
drbrain reason -b "does method A actually solve problem B, or just claim to?"
drbrain reason -b -r 5 "map the causal chain from X to Y across these papers"
```

Options: `-b` / `--bidirectional` (LLM↔KG validation loop), `-r N` / `--max-rounds` (default 3).

## When to use ask vs reason

| Situation | Use |
|-----------|-----|
| Factual lookup ("what is X?") | `ask` |
| Concept comparison ("A vs B") | `ask` (simple) / `reason` (complex) |
| Multi-paper synthesis | `reason` |
| Causal chain tracing | `reason -b` |
| Hypothesis validation | `reason -b` |
| Contradiction resolution | `reason -b` |

## Examples

**Quick concept lookup:**
```bash
drbrain ask "what loss functions are used for contrastive learning?"
```

**Deep multi-paper reasoning:**
```bash
drbrain reason "how do different papers address the over-smoothing problem in GNNs, and which approach is most promising?"
```

**Hypothesis validation with bidirectional loop:**
```bash
drbrain reason -b -r 5 "does the proposed method actually solve the cold-start problem, or does it just shift the burden to feature engineering?"
```

## CLI Reference

| Command | What it does |
|---------|--------------|
| `drbrain ask "<question>"` | KGQA: retrieve → answer |
| `drbrain ask "<q>" --top 10` | More concepts in context |
| `drbrain ask "<q>" --json` | JSON output with provenance |
| `drbrain reason "<question>"` | LLM agent with tool-calling |
| `drbrain reason -b "<q>"` | Bidirectional LLM↔KG validation |
| `drbrain reason -b -r 5 "<q>"` | More hypothesis-revision rounds |

From UynajGI/drbrain, MIT.

More from UynajGI/drbrain