ra-h-os/docs/2_schema.md

Database Schema

Why SQLite?

RA-H uses SQLite for local-first data ownership. Your knowledge stays on your machine - no cloud dependencies. SQLite provides:

• Zero configuration - single file database
• sqlite-vec extension - fast vector similarity search
• Full-text search (FTS5) - Google-like text search
• Relational integrity - foreign keys, triggers, transactions
• Portability - database file migrates with Mac app

Database Location: ~/Library/Application Support/RA-H/db/rah.sqlite

Two-Layer Embedding Architecture

RA-H uses two types of embeddings for different search needs:

1. Node-Level Embeddings

• Storage: nodes.embedding column (BLOB)
• Purpose: Semantic search for nodes (legacy memory pipeline used this too)
• Model: text-embedding-3-small (1536 dimensions)
• Used by: Search/agent tools (legacy memory pipeline has been removed)

2. Chunk-Level Embeddings

• Storage: chunks table (text) → vec_chunks virtual table (embeddings)
• Purpose: Detailed content search within long documents
• Model: text-embedding-3-small (1536 dimensions)
• Used by: searchContentEmbeddings tool

Core Tables

nodes

Primary knowledge storage. Each row is a discrete knowledge item.

Columns:

• id (INTEGER PK) - Unique identifier
• title (TEXT) - Node title
• content (TEXT) - Full content
• type (TEXT) - Node type (memory, paper, idea, person, etc)
• link (TEXT) - External source URL (only for source nodes, not derived ideas)
• description (TEXT) - Brief summary
• metadata (TEXT) - JSON metadata
• chunk (TEXT) - Source text for chunking
• chunk_status (TEXT) - Chunking status (not_chunked, chunked)
• embedding (BLOB) - Node-level embedding vector
• embedding_text (TEXT) - Text that was embedded
• embedding_updated_at (TEXT) - Embedding timestamp
• is_pinned (INTEGER) - Legacy pin flag (kept for migration; not surfaced in UI)
• created_at, updated_at (TEXT) - Timestamps

Indexes:

• idx_nodes_type - Fast type filtering
• idx_nodes_pinned - Legacy partial index (no longer recreated, safe to drop later)

FTS:

• nodes_fts - Full-text search on title + content

edges

Directed relationships between nodes (knowledge graph).

Important behavior:

• Storage is directed (from_node_id → to_node_id)
• UI treats connections as bidirectional (a node shows edges where it is either from or to)
• Every new edge requires an explanation (enforced in service layer)
• Every new edge is classified into a structured EdgeContext (stored as JSON)

Columns (SQLite):

• id (INTEGER PK)
• from_node_id (INTEGER FK → nodes.id) — directed “from”
• to_node_id (INTEGER FK → nodes.id) — directed “to”
• source (TEXT) — creation source (user, helper_name, ai_similarity)
• created_at (TEXT)
• context (TEXT) — JSON blob (canonical; see EdgeContext below)
• explanation (TEXT) — legacy column (currently not the canonical source of truth)
• user_feedback (INTEGER) — user rating (not currently used in core flows)

Indexes:

• idx_edges_from — fast “outgoing edges” queries
• idx_edges_to — fast “incoming edges” queries

EdgeContext (canonical relationship metadata)

Stored as JSON in edges.context. This is the “Idea Genealogy” layer.

interface EdgeContext {
  // SYSTEM-INFERRED (AI + heuristics classify from explanation + node context)
  category: 'attribution' | 'intellectual';
  type:
    | 'created_by'   // attribution: authorship/creation/founding
    | 'features'     // attribution: appears in / host / guest / explicitly mentioned
    | 'part_of'      // attribution: membership/container (episode→podcast, chapter→book, video→channel)
    | 'source_of'    // intellectual: idea/insight came from source
    | 'extends'      // intellectual: builds on
    | 'supports'     // intellectual: evidence for
    | 'contradicts'  // intellectual: in tension
    | 'related_to';  // intellectual: fallback
  confidence: number;   // 0–1
  inferred_at: string;  // ISO timestamp

  // PROVIDED BY USER/AGENT
  explanation: string;  // required; free-form text (user can edit)

  // SYSTEM-MANAGED
  created_via: 'ui' | 'agent' | 'mcp' | 'workflow' | 'quicklink';
}

Direction rule (how to write explanations)

Explanations must read correctly FROM → TO:

• created_by: FROM was created/authored/founded by TO
• features: FROM features/mentions TO
• part_of: FROM is part of TO
• source_of: FROM came from / was inspired by TO

Inference + guardrails

On edge create and on explanation edits, RA-H:

• runs lightweight heuristics for common phrases (e.g., “Created by …”, “Part of …”, “Came from …”, “Features …”)
• otherwise runs an AI classification step to populate category/type/confidence

The UI also provides 4 quick chips to reduce user cognitive load:

• Made by → “Created by …”
• Part of → “Part of …”
• Came from → “Came from …”
• Related → “Related to …”

Where edges get created/updated

All edge creation funnels through the service layer enforcement:

• UI (FocusPanel) — requires explanation; allows editing explanation (re-infers)
• REST API POST /api/edges — requires explanation
• Tooling (createEdge tool) — requires explanation
• MCP (rah_create_edge) — requires explanation
• Workflows (e.g. connect, integrate) — call createEdge with explanation

chunks

Long-form content split into searchable pieces.

Columns:

• id (INTEGER PK)
• node_id (INTEGER FK → nodes.id)
• chunk_idx (INTEGER) - Sequence number
• text (TEXT) - Chunk content
• embedding_type (TEXT) - Model used
• metadata (TEXT) - JSON metadata
• created_at (TEXT)

Indexes:

• idx_chunks_by_node - Fast node→chunks lookup
• idx_chunks_by_node_idx - Ordered retrieval

FTS:

• chunks_fts - Full-text search within chunks

dimensions

Master list of categorization tags.

Columns:

• name (TEXT PK) - Dimension name
• is_priority (INTEGER) - Priority dimension flag
• updated_at (TEXT)

node_dimensions

Many-to-many junction table (nodes ↔ dimensions).

Columns:

• node_id (INTEGER FK → nodes.id)
• dimension (TEXT FK → dimensions.name)
• Primary key: (node_id, dimension)

Indexes:

• idx_dim_by_dimension - Fast "all nodes in dimension X"
• idx_dim_by_node - Fast "all dimensions for node X"

chats

Conversation history with token/cost tracking.

Columns:

• id (INTEGER PK)
• chat_type (TEXT) - Conversation type
• helper_name (TEXT) - Agent key (ra-h, ra-h-easy, mini-rah, wise-rah)
• agent_type (TEXT) - Role category (orchestrator, executor, planner)
• delegation_id (INTEGER FK)
• user_message (TEXT)
• assistant_message (TEXT)
• thread_id (TEXT) - Conversation thread
• focused_node_id (INTEGER FK → nodes.id)
• metadata (TEXT) - JSON with token counts, costs, traces
• created_at (TEXT)

Indexes:

• idx_chats_thread - Fast thread retrieval

agent_delegations

Task queue for mini-rah workers.

Columns:

• id (INTEGER PK)
• session_id (TEXT UNIQUE) - Delegation identifier
• agent_type (TEXT) - Delegate type (default: 'mini')
• task (TEXT) - Task description
• context (TEXT) - Execution context
• expected_outcome (TEXT)
• status (TEXT) - queued, in_progress, completed, failed
• summary (TEXT) - Result summary
• created_at, updated_at (TEXT)

chat_memory_state

Checkpoint tracker for memory pipeline.

Columns:

• thread_id (TEXT PK)
• helper_name (TEXT) - Agent that owns the thread
• last_processed_chat_id (INTEGER) - Last chat processed
• last_processed_at (TEXT)

Indexes:

• idx_chat_memory_thread - Fast state lookup

logs

Activity audit trail (auto-pruned to last 10k).

Columns:

• id (INTEGER PK)
• ts (TEXT) - Timestamp
• table_name (TEXT) - Affected table
• action (TEXT) - INSERT, UPDATE, DELETE
• row_id (INTEGER) - Affected row
• summary (TEXT) - Human-readable summary
• snapshot_json (TEXT) - Row snapshot
• enriched_summary (TEXT) - Enriched log entry

Indexes:

• idx_logs_ts - Chronological queries
• idx_logs_table_ts - Per-table chronological
• idx_logs_table_row - Per-row history
• idx_logs_enriched - Enriched-only filtering

Vector Tables (Auto-Created)

vec_nodes

Virtual table for node-level vector search (sqlite-vec).

VIRTUAL TABLE USING vec0(
  node_id INTEGER PRIMARY KEY,
  embedding FLOAT[1536]
)

Supporting tables (auto-generated):

• vec_nodes_info, vec_nodes_chunks, vec_nodes_rowids, vec_nodes_vector_chunks00

vec_chunks

Virtual table for chunk-level vector search.

VIRTUAL TABLE USING vec0(
  chunk_id INTEGER PRIMARY KEY,
  embedding FLOAT[1536]
)

Supporting tables (auto-generated):

• vec_chunks_info, vec_chunks_chunks, vec_chunks_rowids, vec_chunks_vector_chunks00

Note: Vec tables are NOT pre-seeded in distribution. They auto-create on first app startup via ensureVectorTables() in sqlite-client.ts:45.

Views

nodes_v

Nodes with dimensions aggregated as JSON array.

SELECT 
  n.id, n.title, n.content, n.link, n.metadata, n.chunk,
  n.created_at, n.updated_at,
  COALESCE(JSON_GROUP_ARRAY(d.dimension), '[]') AS dimensions_json
FROM nodes n
LEFT JOIN node_dimensions d ON d.node_id = n.id

logs_v

Enriched logs with related data (node titles, edge titles, chat previews).

Triggers

Logging triggers:

• trg_nodes_ai / trg_nodes_au - Log node inserts/updates
• trg_edges_ai / trg_edges_au - Log edge inserts/updates
• trg_chats_ai - Log chat inserts (with token/cost/trace metadata)

Maintenance triggers:

• trg_edges_update_nodes_on_insert - Touch node timestamps on edge creation
• trg_logs_prune - Keep last 10,000 log rows

Schema Version

schema_version table:

• Tracks database migrations
• Current: v1.0 (frozen for Mac app release)

CREATE TABLE schema_version (
  version INTEGER PRIMARY KEY,
  applied_at TEXT DEFAULT CURRENT_TIMESTAMP,
  description TEXT
);

Seed Database

Location: /dist/resources/rah_seed.sqlite
Purpose: Ships with Mac app for new users
Contents: Clean schema, no data, no vec tables (auto-created on first run)
Size: ~128KB