Files
ra-h-os/docs/4_tools-and-workflows.md
T
“BeeRad” 0f408bc907 docs: sync documentation from private repo
- Rewrite 3_context-and-memory.md (auto-context system, remove legacy memory)
- Create 7_voice.md (voice interface documentation)
- Update 0_overview.md, 1_architecture.md, 4_tools-and-workflows.md, 6_ui.md, 8_mcp.md
- Add simple human-readable intros to all docs
- Update README.md with voice doc reference

Synced from ra-h commit 4a3d7e0
2026-01-05 12:42:51 +11:00

317 lines
7.4 KiB
Markdown

# Tools & Workflows
> What actions the AI can take and how workflows automate multi-step processes.
**How it works:** AI agents have access to tools — functions they can call to read data, create nodes, search content, and more. Workflows are pre-written instruction sets that guide the AI through complex multi-step processes like finding connections across your knowledge base.
---
## Tools Overview
Tools are organized into three categories:
| Category | Purpose | Examples |
|----------|---------|----------|
| **Core** | Read-only graph operations | queryNodes, searchContentEmbeddings |
| **Orchestration** | Delegation and reasoning | executeWorkflow, think, webSearch |
| **Execution** | Write operations and extraction | createNode, updateNode, youtubeExtract |
---
## Core Tools (All Agents)
Read-only operations available to all agents:
### queryNodes
Search nodes by title, content, or dimensions.
```typescript
queryNodes({
search?: string, // Full-text search
dimensions?: string[],// Filter by dimensions
limit?: number // Max results (default: 20)
})
```
### getNodesById
Retrieve full node data by ID array.
```typescript
getNodesById({
ids: number[] // Array of node IDs
})
```
### queryEdge
Inspect existing edges between nodes.
```typescript
queryEdge({
from_node_id?: number,
to_node_id?: number,
limit?: number
})
```
### searchContentEmbeddings
Semantic search across chunk embeddings.
```typescript
searchContentEmbeddings({
query: string, // Search query
node_id?: number, // Scope to specific node
limit?: number, // Max results
threshold?: number // Similarity threshold (0-1)
})
```
### queryDimensions
Query and filter dimensions.
```typescript
queryDimensions({
search?: string, // Filter by name
isPriority?: boolean, // Filter locked dimensions only
limit?: number
})
```
### getDimension
Get a single dimension by exact name.
```typescript
getDimension({
name: string // Exact dimension name
})
```
---
## Orchestration Tools (Orchestrators Only)
Tools for reasoning and delegation:
### webSearch
External web search via Tavily.
```typescript
webSearch({
query: string,
max_results?: number
})
```
### think
Internal reasoning/planning (logged to metadata, not shown to user).
```typescript
think({
thought: string // Reasoning to log
})
```
### executeWorkflow
Delegate to wise-rah for predefined workflows.
```typescript
executeWorkflow({
workflow_key: string // e.g., "integrate"
})
```
---
## Execution Tools (Writers)
Write operations and content extraction:
### createNode
Create a new knowledge node.
```typescript
createNode({
title: string,
content?: string,
description?: string,
dimensions?: string[],
link?: string,
metadata?: object
})
```
### updateNode
Append content to existing nodes. **Append-only** — cannot overwrite.
```typescript
updateNode({
id: number,
content: string // Appended to existing content
})
```
### createEdge
Create relationship between nodes.
```typescript
createEdge({
from_node_id: number,
to_node_id: number,
context?: string // Relationship description
})
```
### updateEdge
Modify edge metadata.
```typescript
updateEdge({
id: number,
context?: string,
user_feedback?: number
})
```
### Dimension Tools
```typescript
createDimension({ name: string, description?: string })
updateDimension({ name: string, description?: string })
lockDimension({ name: string }) // Make priority dimension
unlockDimension({ name: string }) // Remove priority status
deleteDimension({ name: string })
```
### Extraction Tools
```typescript
youtubeExtract({ url: string }) // Extract transcript
websiteExtract({ url: string }) // Extract page content
paperExtract({ url: string }) // Extract PDF text
```
---
## Tool Access by Agent
| Agent | Core | Orchestration | Execution |
|-------|------|---------------|-----------|
| **ra-h / ra-h-easy** | ✅ All | webSearch, think, executeWorkflow | ✅ All |
| **wise-rah** | ✅ All | webSearch, think, delegateToMiniRAH | updateNode, createEdge |
| **mini-rah** | ✅ All | webSearch, think | ✅ All |
---
## Workflows
Workflows are multi-step instruction sets executed by wise-rah.
### User-Editable Workflows
**Users can create, edit, and delete workflows** from Settings → Workflows tab.
**Storage:** `~/Library/Application Support/RA-H/workflows/`
**Format:** JSON files with this structure:
```json
{
"key": "integrate",
"displayName": "Integrate",
"description": "Deep analysis and connection-building for focused node",
"instructions": "You are executing the INTEGRATE workflow...",
"enabled": true,
"requiresFocusedNode": true
}
```
### How Workflows Work
1. User says "run integrate workflow" (or similar)
2. Orchestrator calls `executeWorkflow({ workflow_key: 'integrate' })`
3. System spawns wise-rah session with workflow instructions
4. wise-rah executes steps autonomously (30-60+ seconds)
5. wise-rah returns summary to orchestrator
6. Orchestrator shows result to user
### Built-in Workflows
#### Integrate
**Key:** `integrate`
**Purpose:** Database-wide connection discovery
**Cost:** ~$0.18/execution
**5-Step Process:**
1. **Plan** — Call `think` to outline approach
2. **Ground** — Extract entities (names, projects, concepts) from focused node
3. **Search** — Database-wide search using extracted entities
4. **Contextualize** — Brief relevance explanation
5. **Append** — Call `updateNode` ONCE with Integration Analysis section
**Output Format:**
```markdown
## Integration Analysis
[2-3 sentences: what this node is, why it matters]
**Database Connections:**
- [NODE:123:"Title"] — [why relevant]
- [NODE:456:"Title"] — [why relevant]
...
**Relevance:** [1-2 sentences connecting to your context]
```
### Creating Custom Workflows
1. Go to Settings → Workflows
2. Click "New Workflow"
3. Fill in:
- **Key** — unique identifier (lowercase, no spaces)
- **Display Name** — shown in UI
- **Description** — what it does
- **Instructions** — the prompt for wise-rah
- **Requires Focused Node** — whether a node must be selected
4. Save
### Workflow Instructions Tips
- Be explicit about steps and expected output
- Reference tools by name (wise-rah has: queryNodes, searchContentEmbeddings, updateNode, createEdge, webSearch, think)
- Specify output format precisely
- Include guardrails ("call updateNode exactly once")
---
## Tool Registry
**Location:** `src/tools/infrastructure/registry.ts`
**Structure:**
```typescript
TOOL_SETS = {
core: { queryNodes, getNodesById, queryEdge, queryDimensions, getDimension, searchContentEmbeddings },
orchestration: { webSearch, think, delegateToMiniRAH, executeWorkflow, ... },
execution: { createNode, updateNode, createEdge, updateEdge, youtubeExtract, websiteExtract, paperExtract, ... }
}
```
---
## Key Design Decisions
### Append-Only Updates
`updateNode` is **append-only** — it cannot overwrite existing content. This prevents AI from accidentally destroying knowledge. The tool-level enforcement means workflows can't bypass this restriction.
### No Workflow Delegation Loops
wise-rah cannot call `executeWorkflow` or delegate to other wise-rah instances. This prevents infinite loops and keeps execution bounded.
### Isolated Execution
Workers (wise-rah, mini-rah) execute in isolated sessions. They return structured summaries only — they don't pollute orchestrator context with execution details.