Thomashighbaugh/vectorize-context

Vectorize Context

Indexes .opencode/context/ markdown files into a local sqlite-vec vector database for semantic search. Manual trigger only — invoked via /harvest-context search.

How It Works

The system uses lazy freshness: every query automatically stats all context files, re-indexes only what changed, then searches. This covers every write path:

| Trigger | Behavior | |---------|----------| | Hub writes a new context/decision/pattern file | Next query picks it up automatically | | /harvest-context saves research docs | Indexed on next query | | /orchestrate completes and saves patterns | Indexed on next query | | /ideation finalizes a plan to context | Indexed on next query | | Direct .md file edit in .opencode/context/ | Indexed on next query | | Full re-index needed | Delete .opencode/.vector/ — auto-rebuilds |

The ML model is only loaded when there's actual work to do (files changed). If everything is up to date, ensureIndexed() returns instantly with no model load.

Scripts

| Script | Purpose | |--------|---------| | scripts/veclib.mjs | Shared library — exported programmatic API for all integrations | | scripts/vectorize.mjs | CLI: manual re-index (for debugging/forced refresh) | | scripts/query.mjs | CLI: semantic search over context |

Programmatic API (veclib.mjs)

Used by Hubs hub subcommands and agents:

// Lazy re-index: stats files, only loads model if something changed
import { ensureIndexed } from './veclib.mjs';
await ensureIndexed();                // current project
await ensureIndexed('/path/to/app');  // specific project

// Semantic search (auto-refreshes index first)
import { queryChunks } from './veclib.mjs';
const results = await queryChunks(undefined, 'auth patterns', 10);
// results: [{ source, heading, content, file_path, distance }]

// Index stats
import { getIndexStats } from './veclib.mjs';
const stats = await getIndexStats();
// { exists, totalChunks, totalFiles, files: [...] }

CLI Usage

# Manual re-index (useful for debugging)
node {skill_dir}/scripts/vectorize.mjs

# Semantic search (auto-refreshes on every call)
node {skill_dir}/scripts/query.mjs "how does error handling work"
QUERY="auth patterns" node {skill_dir}/scripts/query.mjs

Output Format

Query results

=== Search Results ===

1. patterns/error-handling.md — Error Patterns (score: 0.71)
   Error handling follows a centralized approach with...
   [file: .opencode/context/patterns/error-handling.md]

2. frameworks/architecture.md — System Design (score: 0.64)
   The project follows a layered architecture...
   [file: .opencode/context/frameworks/architecture.md]

Dependencies

npm install better-sqlite3 sqlite-vec @xenova/transformers

Requires Node.js 18+ and ~200MB RAM for the embedding model (loaded lazily only when indexing).

Integration Points

Hub Lifecycle Integration (manual trigger)

After any hub subcommand writes to .opencode/context/:

1. Write the file (existing behavior)
2. Manual vectorize: Run /harvest-context search to index and query — no automatic indexing

Agent Integration

Agents can use queryChunks() to retrieve relevant context during execution:

const ctx = await queryChunks(process.cwd() + '/.opencode', query, 5);
// Inject results into agent context as supporting evidence

Forced Re-index

rm -rf .opencode/.vector/    # Delete the DB
# Next query or ensureIndexed() call auto-rebuilds it

How It Works Internally

1. ensureIndexed() scans .opencode/context/ for *.md files
2. Compares current file mtimes against stored mtimes in the DB
3. If no files changed → returns immediately, no model loaded
4. If files changed → loads model, chunks new/changed files by ##/### headers
5. Generates 384-dim embeddings via Xenova/all-MiniLM-L6-v2
6. Deletes old chunks for changed files, inserts new ones
7. Query always runs against the freshly-updated index

The vec0 virtual table uses L2 distance. Since embeddings are normalized (unit vectors), L2 distance sorts equivalently to cosine similarity — nearest neighbors are the most semantically similar chunks.