seeklink 0.3.2

Hybrid semantic search CLI for Obsidian-compatible markdown vaults. BM25 + vector + knowledge graph + title, fused via RRF. Native CJK. Fully local.

SeekLink

Hybrid semantic search for your Obsidian-compatible markdown vault.

SeekLink searches your personal knowledge base using four channels in parallel — keyword matching, semantic similarity, knowledge graph, and title/alias lookup — then fuses the results for high-recall, high-precision retrieval. Optional cross-encoder reranking via MLX gives an extra precision boost on Apple Silicon.

Built for people who take notes seriously and want an AI that understands their knowledge structure, not just their text.

Ideal for: Obsidian power users, Zettelkasten / second-brain practitioners, bilingual English + Chinese note-takers, and anyone building local-first RAG on top of a markdown vault. Works with any folder of .md files — no Obsidian plugin required.

What it does

You:   seeklink search "agent memory systems" --vault ~/notes
       → 8 related notes across topics, ranked by relevance

You:   seeklink search "记忆保持力" --vault ~/notes --title-weight 0.5
       → surfaces raw log entries alongside polished articles

You:   seeklink daemon --vault ~/notes
       → resident mode: first search ~2s; warm reranker-on ~1-2s,
         reranker-disabled ~10ms

When to use SeekLink

• Your knowledge base is a folder of markdown files, with or without Obsidian.
• You want search that understands meaning, not just exact words — and that surfaces short log entries on equal footing with titled permanent notes.
• You want it to work offline, with no API keys, no cloud, and no data leaving your machine.
• You write in English, Chinese, or both. CJK is a first-class tokenization path (jieba as a custom FTS5 tokenizer), not an afterthought.
• You want a CLI an agent can shell out to. Any tool that can exec() a binary can use SeekLink — no MCP client or vendor lock-in.

When not to use SeekLink

• Your notes are not markdown (Notion export, Bear, Apple Notes native format, Roam). Convert them first, or pick a tool built for that format.
• You want a hosted, synced, multi-user search service. SeekLink is single-machine, single-user.
• You want a GUI inside Obsidian. SeekLink is a CLI + daemon — there's no Obsidian plugin (yet).
• You need sub-millisecond search over millions of notes. SeekLink targets personal vaults (thousands to low tens-of-thousands of notes). At that scale it's fast; beyond, you want a real search service (Typesense, Meilisearch, Elastic).
• You're on Windows. macOS and Linux are tested; Windows should mostly work (Python is portable, the Unix-socket daemon isn't) but is not a supported path.

How it compares

Tool	Format	Semantic search	CJK	Local-only	Knowledge-graph signal
SeekLink	any .md vault	✅ BM25 + vector + title, RRF-fused, optional MLX reranker	✅ jieba tokenizer	✅	✅ wikilink indegree
Obsidian core search	Obsidian vault	❌ (keyword only)	partial	✅	❌
Obsidian semantic-search plugins (Smart Connections, Copilot, etc.)	Obsidian vault	✅ (vector only, mostly OpenAI API)	varies	usually ❌ (API)	❌
ripgrep	any text	❌ (keyword only, literal regex)	✅	✅	❌
Build-your-own RAG (LlamaIndex, LangChain)	any	✅ (you wire it)	depends on embedder	depends	depends

SeekLink's niche: markdown-native, hybrid keyword + semantic, CJK-first, fully local, with the knowledge-graph signal your wikilinks already encode — available as a single pip install.

Architecture

Query: "agent memory systems"
        │
        ├── BM25 (FTS5 + jieba) ──── keyword match ──────── weight 1.0
        ├── Vector (jina-v2-zh) ──── semantic similarity ── weight 1.0
        ├── Indegree ─────────────── well-linked = quality ─ weight 0.3
        └── Title/Alias (FTS5) ──── exact name match ────── weight 1.5
        │
        └── RRF Fusion → top candidates
                │
                └── [optional] Qwen3-Reranker-0.6B (MLX) → precision boost
                        │
                        └── ranked results

Four-channel Reciprocal Rank Fusion, with optional cross-encoder reranking for Apple Silicon. Everything runs locally — no API keys, no cloud.

Support & limitations

	Supported	Notes
Python	3.11, 3.12, 3.13, 3.14	Tested in CI
OS (core + daemon)	macOS, Linux	Unix socket at ~/.rhizome/seeklink.sock
OS (Windows)	unsupported	Python code is portable; Unix-socket daemon is not
File format	.md (markdown)	Frontmatter optional
Wikilink syntax	[[note]], [[alias]], Obsidian-compatible
Embedder	jina-embeddings-v2-base-zh (default, 330 MB)	Swap via SEEKLINK_EMBEDDER_MODEL (fastembed-supported)
Reranker	Qwen3-Reranker-0.6B via MLX (700 MB)	Apple Silicon only. Disable with SEEKLINK_RERANKER_MODEL=""
Multi-vault daemon	single-vault only	Pass --vault to run a one-shot against a different vault (forces cold-start)
Concurrent access	one daemon per machine	Multiple CLI clients may share one daemon

Install

uv tool install seeklink
# or
pip install seeklink

Quick start

# 1. Build the index first (searching an un-indexed vault returns no hits).
seeklink index --vault /path/to/vault

# 2. Search it.
seeklink search "machine learning" --vault /path/to/vault

Recommended for daily use — set your default vault once, then every search / index auto-uses the fast daemon:

export SEEKLINK_VAULT=/path/to/vault
seeklink index                         # first run: builds the index
seeklink search "machine learning"
# First search after a cold boot: ~2s (spawns the daemon, loads the embedder).
# Warm reranker-on path (default): ~1-2s per query.
# Warm reranker-disabled path: ~10ms per query.

The daemon stays resident across terminal sessions until you kill it or restart. seeklink search and seeklink index auto-spawn it when missing; seeklink status is always cold-start (it only reads SQLite stats, no model load) and seeklink get is a direct filesystem read (no daemon involved either).

For agents

SeekLink is designed to be shelled out to by AI agents (Claude Code, Cursor, custom RAG pipelines), not just typed by humans. The contract:

1. Index once, then search. Agents should check seeklink status --vault PATH before first use; if Notes: 0 or stale, run seeklink index --vault PATH.
2. Search. seeklink search "query" --vault PATH --top-k N. Each result line is:

     SCORE  PATH[:LINE]  TITLE
              <indented content preview, up to 120 chars, one line>
   

   PATH is relative to the vault root. :LINE is 1-indexed and points at the best chunk's first line in the current on-disk file; omitted when the match is title-only on a stale file. SCORE is rerank_blended when reranker is active, else raw RRF (scales differ, don't compare across configurations).
3. Read the window. seeklink get PATH:LINE -l N prints N lines starting at LINE. No DB lookup, no daemon round-trip — direct filesystem read with universal-newline translation. This is how you avoid slurping whole files after a search hit.
4. Query shape. Raw CJK is fine (jieba handles segmentation). [[alias]] / exact-title queries get title-gated protection — if your query IS a note title or alias, results anchor on it. No need for wildcards; BM25 + vector + title fuse automatically.
5. Exit codes. 0 on success (including "no results"). 1 on vault-resolution error, missing file (get), or unrecoverable config mismatch. No other codes used.
6. Structured output. If you want JSON instead of the human text format, connect to the daemon's Unix socket at ~/.rhizome/seeklink.sock directly (length-prefixed JSON protocol, see seeklink/daemon.py). Each daemon search response carries path, title, content_preview, score, indegree, line_start, line_end per result.

See also llms.txt at the repo root for a compressed version of this contract.

CLI reference

seeklink search

seeklink search "query" --vault PATH [options]

Options:
  --top-k N          Number of results (default: 10)
  --tags TAG [TAG]   Filter by tags (AND semantics)
  --folder PREFIX    Filter by folder (e.g. "notes/")
  --title-weight F   Override title channel weight (default: 1.5)
                     Raise toward 3.0 for "find the article" queries;
                     lower toward 0.5 for "surface raw moments" queries.

seeklink daemon

Starts a Unix-socket daemon that keeps the embedding model (and reranker, if enabled) resident in memory. First query after startup takes ~2s (model warmup); subsequent queries return in ~10ms without reranker or ~2s with reranker.

You almost never run this directly. seeklink search and seeklink index auto-spawn a daemon on cold machines when --vault is not passed. seeklink status is always cold-start (no model load). seeklink get is a direct filesystem read (no daemon). The daemon uses SEEKLINK_VAULT (or cwd) as its vault and never auto-exits — kill it with kill or restart your machine.

Passing --vault always uses cold-start instead of the daemon, because the daemon binds to a single vault at startup. Multi-vault daemon support is tracked in TODOS.md.

seeklink daemon --vault PATH    # foreground, for debugging

seeklink index

seeklink index [PATH] --vault VAULT

Without PATH: full vault re-index (detects unchanged files via content hash).
With PATH:    index a single file.

seeklink status

seeklink status --vault PATH

Shows index stats and freshness warnings. If files have changed since last index, prints a warning to stderr.

seeklink get

Print a line range of a vault file directly to stdout. Designed for agents that have a search hit like notes/fsrs.md:42 and want to read a precise window without fetching the whole file.

seeklink get PATH[:LINE] [-l N] [--vault PATH]

  seeklink get notes/fsrs.md              # entire file
  seeklink get notes/fsrs.md:120          # 100 lines starting at line 120
  seeklink get notes/fsrs.md:120 -l 40    # 40 lines starting at line 120
  seeklink get notes/fsrs.md -l 50        # first 50 lines

Line numbers match search output. CRLF files print with universal newlines. Path escapes (../..) are rejected.

How search works

SeekLink runs four search channels in parallel and merges results with Reciprocal Rank Fusion:

• BM25 (FTS5 + jieba): keyword match on chunk content. Handles CJK natively via jieba tokenization.
• Vector (jina-embeddings-v2-base-zh): semantic similarity. Finds conceptually related notes even when they use different words or languages.
• Indegree: notes that many other notes link to rank higher — a lightweight quality signal from your knowledge graph.
• Title/Alias (FTS5): matches against note titles and aliases frontmatter. Weight 1.5 gives a modest boost without overwhelming content matches.

Why title weight is 1.5 (not higher)

Many personal knowledge bases contain a mix of titled articles (permanent notes, literature reviews) and untitled process notes (daily logs, journal entries, quick captures). A high title weight systematically buries untitled content — even when it's the most relevant result for the query. The default of 1.5 keeps title matching useful for precise [[alias]] lookups while letting content-based matches compete on their own merits. Override with --title-weight per query if needed.

Title-gated rerank blending (v0.3+)

When the reranker is enabled, a cross-encoder (Qwen3-Reranker-0.6B on MLX, ~1-2s per query) re-scores the top-20 RRF candidates for precision. SeekLink applies title-gated position blending on top of this:

• If the title channel's best match is in the candidate pool, blend alpha · normalized_rrf + (1 - alpha) · rerank_score with alpha = 0.60/0.50/0.40 by rank bucket. This protects exact title / alias hits from being demoted by a content-focused reranker.
• Otherwise (no strong title signal), the reranker score is used directly — same as pre-v0.3 behavior. This lets the reranker correct poor first-stage ordering.

On the bundled 22-query pilot (see tests/blind/), mean MRR moved from 0.932 to 0.977 vs pure-reranker-override with no per-query regressions. Sample size is a pilot, not a statistically powered benchmark — contributions of larger labeled corpora are welcome.

Disable reranking entirely with: export SEEKLINK_RERANKER_MODEL=""

Results carry line numbers

Every search result is anchored to a specific line range in the current on-disk file. Two surfaces expose this:

• CLI text output (seeklink search ...): each line is SCORE PATH:LINE_START TITLE, followed by an indented content preview. Feed PATH:LINE_START straight into seeklink get to read the window around the hit.
• Daemon JSON (one request per search via the Unix socket): each result also carries line_end and indegree fields for callers that want the full span.

Line numbers are mapped back through the frontmatter strip that happens at index time, so they match what you'd see in a text editor on disk.

Frontmatter

SeekLink works with any markdown file — no special formatting required.

If your notes have YAML frontmatter, SeekLink uses it for extra features:

---
tags: [ai, machine-learning]
aliases: [ML, Machine Learning]
---

• Tags enable filtered search: seeklink search "query" --tags ai
• Aliases are searchable and used for wikilink resolution — [[ML]] resolves to the note with that alias

How it stores data

Everything lives in .seeklink/seeklink.db inside your vault — a single SQLite database with:

• FTS5 full-text index (jieba-tokenized for CJK)
• sqlite-vec for 768-dim vector similarity search
• A wikilink graph (parsed from [[links]] in your notes)

Notes are chunked (~400 tokens), embedded with jina-embeddings-v2-base-zh, and indexed incrementally. Delete .seeklink/ to rebuild from scratch.

Configuration

Variable	Default	Description
SEEKLINK_VAULT	.	Path to vault root
SEEKLINK_EMBEDDER_MODEL	jinaai/jina-embeddings-v2-base-zh	Embedding model (fastembed-supported)
SEEKLINK_RERANKER_MODEL	mlx-community/Qwen3-Reranker-0.6B-mxfp8	Reranker model (set to "" to disable)

What changed in v0.3

• Title-gated rerank blending: when an exact title / alias hit drives rank 1, protect it from reranker demotion; otherwise fall back to pure reranker. Measured mean MRR 0.932 → 0.977 on the bundled 22-query pilot (see "How search works" for caveats on sample size).
• Line-range retrieval: search results now include line_start / line_end, and a new seeklink get PATH[:LINE] -l N command prints line-precise windows. Agents can find-then-read without slurping whole files.
• Cold-start / daemon parity fix: cold-start seeklink search now constructs a Reranker() and passes it to the search pipeline. Previously the same query returned different rankings depending on whether the daemon was running.
• Frontmatter-aware line mapping: chunk offsets (stored against frontmatter-stripped body) are remapped to full-file line numbers, so search + get report lines the way you'd see them in a text editor.
• Blind-test framework at tests/blind/: 32-file corpus + 22 ground-truth queries + runner that measures Recall@10 / MRR / latency. Used to validate v0.3 before tagging; gates v0.4 (query expansion) the same way.

What changed in v0.2

• CLI-first: MCP server removed. All interaction via seeklink search/index/status/daemon.
• Daemon mode: Unix-socket resident server with auto-spawn. Models stay loaded for fast queries.
• Reranker: Qwen3-Reranker-0.6B via MLX on Apple Silicon. Optional, default enabled.
• Freshness check: bidirectional mtime scan replaces the file watcher. Warns on stale/new/deleted files.
• Title weight 1.5: down from 3.0, so log entries and journal notes compete fairly with titled permanent notes.
• Leaner deps: mcp and watchfiles removed. 4 runtime dependencies instead of 6.

Contributing

git clone https://github.com/simonsysun/seeklink
cd seeklink
uv sync --dev
uv run python -m pytest tests/ -q

License

MIT