Release v2.0.0 · arpitnath/blink-query

[2.0.0] - 2026-04-08

npm run benchmark — single-command reproducer that auto-clones three public markdown corpora (Quartz docs, Obsidian Help vault, MDN content) and runs blink-query against grep and ripgrep on the identical file set with verifiable per-query oracles
benchmark/ directory at the repo root: bench.ts single-corpus runner, setup.ts auto-clone, run-all.ts orchestrator with cross-corpus comparison and ASCII speedup bars, README, gitignored corpora and runtime artifacts
Headline numbers from the universal config (one library configuration, no per-corpus tuning): blink-query is 25×–91× faster than grep and 21×–31× faster than ripgrep across the three corpora; P@1 64% / P@3 79% average

Title-weighted BM25 with type-aware rank offsets (searchByKeywords) — bm25(records_fts, 10.0, 4.0, 1.0): title 10×, tags 4×, summary 1×, plus per-record-type rank offsets (SUMMARY −2.0, META −1.0, COLLECTION −0.5). Universal across markdown corpus shapes.
Hub-vs-leaf detection (loadDirectoryBasic) — pre-walk pass tags every document with is_canonical (file is index/readme/home/about of its dir) and is_hub (canonical AND parent dir contains other subdirs). Additive metadata, zero breaking changes.
defaultClassify (new exported function in src/ingest.ts) — promotes canonical hub pages to SUMMARY automatically using the structural signals from the walker. Frontmatter type: field still wins where present. Replaces the previous "everything is SOURCE" default when no classify callback is supplied.
filesystemTitle improvements — for index.md / README.md / Home.md / About.md files, uses the parent directory name as the title. For date-style (Logseq) and UUID-style (Notion export) filenames, falls back to the first H1 in the document body. Behavior change: a previously-saved docs/readme.md now resolves to docs/docs instead of docs/readme.
shortId() 8 → 16 hex chars — fixes UNIQUE constraint failed: records.id at ~14k records (32-bit IDs collided via birthday paradox at ~2.4% probability). 64-bit IDs push the threshold to ~4 billion records.

docs/USING_IN_PROJECT.md — step-by-step library walkthrough: install, hello world, the five record types, five ingest patterns (folder/GitHub/Postgres/git/custom), querying (resolve/search/query/list), mutations, zones, where the data lives, common errors
docs/CONNECTING_TO_AGENT.md — agent setup walkthrough: blink init flow, per-agent walkthroughs for Claude Desktop / Claude Code / Cursor / Codex, verification, dropping BLINK_WIKI.md into agent context, troubleshooting (nvm, Windows, JSON merge, empty results, doctor failures)

WIKI_DERIVERS preset for ingesting markdown sources with namespace, title, type, and tag derivers
[[wikilink]] extraction with auto ALIAS record creation on ingest
blink_ingest MCP tool for bulk ingest from agent sessions
Fix: content field now passes through for all record types (was gated to SOURCE only)

Blink.createZone({ namespace, description, defaultTtl, requiredTags }) — register a zone with metadata. Records saved into the zone inherit the default TTL and must carry the required tags.
Blink.getZone(namespace) — fetch a zone by namespace
blink_create_zone MCP tool — lets agents carve out new namespaces for their workflows with enforced policies
Migration: adds required_tags column to the zones table (idempotent, safe for existing databases)
createWikiNamespace(patterns) factory — extend WIKI_DERIVERS with custom top-level directories (e.g. decisions/, adr/, people/) using a simple template syntax ({dir}, {slug(dir)}). Paths that don't match fall back to the default wiki routing.

blink init — auto-detects agent environment (Claude Desktop, Claude Code, Cursor, Codex) and writes MCP config
blink doctor — post-install diagnostic: checks MCP config, database path, server connectivity
blink wiki init/ingest/lint — subcommand group for wiki workflows

examples/llm-wiki/ — end-to-end wiki on a 30-file MCP ecosystem corpus
Retrieval comparison (examples/llm-wiki/benchmark/): grep vs blink BM25 over the same corpus, no external dependencies

First README.md — "typed wiki for LLMs" framing, benchmark section, quick start, install guide
BLINK_WIKI.md — schema document for LLM agents: namespace conventions, 5 record types, ingest/query/log/lint workflows, 4 worked example sessions

524 tests across 24 suites (up from 388 across 17 in v1.1.0)
New suites: wiki-derivers, wikilink-extraction, install, cli-wiki, zones, create-wiki-namespace, frontmatter
Extended: mcp.test.ts with blink_ingest coverage
vitest config excludes .local/** and benchmark/** from test discovery