blink-query

A typed wiki for LLMs — markdown on disk, resolution in the library.

blink-query is a knowledge resolution library for LLM agents. Markdown files stay on disk where you can read and edit them; the library adds typed records, title-weighted BM25 search, deterministic path resolution, and an MCP server — all in a SQLite file next to your notes. No cloud, no embeddings, no daemon.

Features

• 5 typed record types — SUMMARY, META, SOURCE, ALIAS, COLLECTION — each carries a consumption instruction for the agent (read the summary / parse JSON / fetch the source / follow the redirect / browse children)
• Namespaces — slash-delimited paths group records by topic (e.g., wiki/concepts, decisions/auth, people/arpit); each record has a stable path at namespace/slug-of-title
• Title-weighted BM25 over FTS5 — title matches ranked 10× over body summary, with per-type rank offsets that promote canonical pages
• Hub-vs-leaf default classifier — automatically promotes index / README / Home files in hub directories to SUMMARY, no per-corpus tuning
• Path resolution — deterministic O(1) lookup by namespace/slug, with NXDOMAIN / STALE / ALIAS_LOOP statuses
• MCP server with 11 tools — drop-in for Claude Desktop, Claude Code, Cursor, Codex
• Query DSL — Peggy-parsed: wiki where type = "SUMMARY" and tags contains "auth" limit 10
• Multi-source ingestion — markdown directories, PostgreSQL rows, GitHub Issues, git repos, web URLs
• Zones — top-level namespaces with default TTL and required-tag policies
• [[wikilink]] extraction — auto-creates ALIAS records on ingest

Quick start

npm install -g blink-query
blink init                 # auto-detect agent (Claude Desktop, Code, Cursor, Codex) and write MCP config
blink wiki init my-wiki    # scaffold + ingest a wiki directory
blink doctor               # verify install

Or use it as a library:

npm install blink-query

import { Blink } from 'blink-query';

const blink = new Blink({ dbPath: './wiki.db' });
await blink.ingestDirectory('./my-wiki');
const results = blink.search('how to configure auth');

Step-by-step walkthroughs:

• Using blink-query in your project — install, ingest, query, common patterns, custom derivers, troubleshooting
• Connecting to your agent — Claude Desktop / Code / Cursor / Codex per-agent setup, verification, troubleshooting

Benchmark

One command. Three public markdown corpora. One library configuration.

npm run benchmark

blink-query is 25x–91x faster than grep across 3 corpora, 21x–31x faster than ripgrep. Speed gap widens with corpus size — blink scales sub-linearly, grep scales linearly.

On the 14k-file MDN corpus, grep returns an average of 1,212 unranked files per query (because common terms like "Promise" appear in 1,314 files, "DOM" in 9,363). blink returns top-5 ranked. The agent reads ~242x fewer files to find the answer.

Accuracy across the 3 corpora: P@3 = 79% average (88% on Obsidian, 79% on Quartz, 72% on MDN). Verifiable regex oracles for every query in benchmark/bench.ts. Methodology, caveats, and reference numbers in benchmark/README.md.

The bench script auto-clones the corpora into benchmark/corpora/ (gitignored, ~270 MB on first run), runs blink + grep + ripgrep on the same file set with the same query set, and prints a unified comparison. No --mode flag. No per-corpus tuning. No presets.

What this is, and isn't

blink-query is a library, not a framework. It's additive on top of plain markdown:

• Your files stay where they are. Markdown in sources/, entity/, topics/, log/<date>/. You can grep them, open them in any editor, commit them to git, nothing about them is opaque.
• The library adds an index, types, and a query interface. Think of blink-query as a deterministic faster path on top of the same files.
• No embeddings. BM25 / FTS5 only. If you want semantic retrieval, run a vector store alongside and pick whichever answer is better for the query.
• No cloud. SQLite file on disk. MCP server over stdio. Runs entirely on your machine.
• No magic. Ingestion is rule-based derivers you can override. Classification is based on filename + frontmatter you control. The query DSL is a Peggy grammar you can read.

If you want raw markdown + grep, the files still work that way. If you want typed records with an index, install the library.

Five record types

The type is a consumption instruction — it tells the agent how to use the record, not what it's about. Content carries the domain semantics.

SUMMARY — read the summary directly, you have what you need.
Use for processed wiki pages, entity descriptions, topic overviews. The agent doesn't need to fetch anything else.

META — structured data (JSON content field).
Use for configuration, log entries, session state, entity attributes. The agent parses content as JSON.

SOURCE — summary here, fetch the source if you need depth.
Use for references to external documents: papers, URLs, git files, API specs. Summary gives enough to decide whether to fetch.

ALIAS — follow the redirect to the target record.
Use for cross-linking. [[wikilinks]] in your markdown auto-extract to ALIAS records on ingest.

COLLECTION — browse children, pick what's relevant.
Use for namespace indexes. Auto-generated when you resolve a namespace path that has no direct record.

Install in your agent

The easy path:

npx blink-query init

This auto-detects Claude Desktop, Claude Code, Cursor, and Codex. It handles the nvm + absolute-node-path gotcha on macOS/Linux, wraps npx in cmd /c on Windows, and merges into your existing MCP config — it never overwrites other servers (if the JSON fails to parse it backs up to .bak and warns).

Run blink doctor afterwards to verify.

Manual config

Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "blink": {
      "command": "npx",
      "args": ["-y", "blink-query", "mcp"],
      "env": { "BLINK_DB_PATH": "~/.blink/blink.db" }
    }
  }
}

Claude Code:

claude mcp add-json blink --scope user '{"type":"stdio","command":"npx","args":["-y","blink-query","mcp"]}'

Cursor — ~/.cursor/mcp.json (same shape as Claude Desktop, key mcpServers).

Codex — ~/.codex/config.toml:

[mcp_servers.blink]
command = "npx"
args = ["-y", "blink-query", "mcp"]

[mcp_servers.blink.env]
BLINK_DB_PATH = "~/.blink/blink.db"

nvm users: if your agent can't find npx, run blink init --absolute-node to write the absolute path to your node binary instead. Avoids the "agent doesn't source shell rc" issue.

Runtime capabilities

Once connected, the agent can resolve, search, save new records, move or delete existing ones, create new zones with required-tag policies, and ingest fresh markdown — all via MCP tool calls at runtime. No restart, no re-deploy, no re-ingest. The knowledge base grows under the agent's own hands.

The MCP server exposes 11 tools, grouped by what they do:

Read

Write

Zones

Drop BLINK_WIKI.md into your project root (or paste it into your agent's system prompt) so the agent knows how to use the wiki pattern.

CLI

# Wiki workflows
blink wiki init my-wiki          # scaffold + ingest
blink wiki ingest ./my-wiki      # re-ingest after changes
blink wiki lint                  # find STALE, NXDOMAIN, broken aliases

# Core operations
blink resolve wiki/mcp-protocol  # deterministic O(1) lookup
blink search "tool call protocol"
blink query 'wiki where type = "SUMMARY" limit 10'
blink list wiki --limit 20

# Agent setup
blink init                       # write MCP config for detected agent
blink doctor                     # post-install diagnostic

# Data management
blink ingest ./docs --prefix wiki
blink move wiki/old wiki/new
blink delete wiki/outdated
blink zones

All commands support --json for machine-readable output and --db for a custom database path.

Library API

import { Blink, extractiveSummarize } from 'blink-query';

const blink = new Blink({ dbPath: './wiki.db' });

// Ingest a directory of markdown files
await blink.ingestDirectory('./my-wiki', {
  summarize: extractiveSummarize(500),
  namespacePrefix: 'wiki',
});

// Resolve — deterministic O(1) path lookup
const response = blink.resolve('wiki/mcp-protocol');
if (response.status === 'OK') console.log(response.record.summary);

// Search — title-weighted BM25 over typed records
const results = blink.search('tool call protocol');

// Query DSL
const summaries = blink.query(
  'wiki where type = "SUMMARY" order by hit_count desc limit 10',
);

// Save a new record (idempotent on path — upserts by namespace + slug-of-title)
blink.save({
  namespace: 'wiki/concepts',
  title: 'OAuth Flow',
  type: 'SUMMARY',
  summary: 'OAuth 2.1 flow for first-party apps — PKCE, refresh tokens, session scoping.',
  tags: ['auth'],
});

// Create a zone with a required-tag policy
blink.createZone({
  namespace: 'decisions',
  description: 'Architecture decision records',
  defaultTtl: 31536000,    // 1 year, inherited by saves into this zone
  requiredTags: ['adr'],   // save() throws if missing
});

blink.close();

All CRUD operations are synchronous (resolve, get, save, delete, move, search, list, query, createZone, zones). Only ingestion is async.

Schema doc for your agent

BLINK_WIKI.md is the document your LLM agent reads to understand how to use blink-query as a wiki. It covers namespace conventions, all five record types with examples, ingest/query/log/lint workflows, and four worked example sessions.

Drop it in your project root or paste it into your agent's system prompt.

Example

examples/llm-wiki/ is a complete end-to-end example: a 30-file MCP ecosystem corpus ingested, queried, and benchmarked against grep.

Scope

blink-query handles:

• Markdown and plain text (primary)
• PostgreSQL tables (via loadFromPostgres)
• Web URLs (via loadFromURL)
• Git repositories (via loadFromGit)
• GitHub Issues (via loadFromGitHubIssues)

It does not do:

• Vector embeddings — BM25/FTS5 only. Layer a RAG system on top if you need semantic similarity.
• Sync or replication — single SQLite file, local. Sync the markdown source via git and re-ingest.
• Hosted storage — no cloud, no accounts. Your data lives in a .db file you own.
• Agent orchestration — blink-query resolves knowledge. It doesn't plan tasks or coordinate agents.

Development

npm install
npm run build       # builds parser + library + CLI
npm test            # 524 tests
npm run benchmark   # universal benchmark on 3 public corpora

Status

v2.0.0 — published on npm as blink-query. MIT license.

npm install blink-query        # library
npm install -g blink-query     # CLI + blink init

Issues and PRs welcome.