dshbase

Plugin directory / Knowledge / semantic-search

semantic-search

Verified · install-tested on dsh JohnXu22786

✓ Actively maintained Builds on 2 official DSH packages

View on GitHub ↗ ← Back to plugin directory

0Stars
0Forks
0Open issues
—Language
2026-08-20Last push
Cross-platformPlatform

What it does

Local semantic code search for DeepSeek Harness (dsh): fragment-level, symbol-aware indexing, offline lexical embeddings or OpenAI-compatible endpoints, and hybrid vector + BM25 retrieval fused with RRF - three dsh tools (sema_search / sema_reindex / sema_stats) plus a sema CLI.

✅
Our take
Works — verified, early-stage project

Local semantic code search for DeepSeek Harness (dsh): fragment-level, symbol-aware indexing, offline lexical embeddings or OpenAI-compatible endpoints, and hybrid vector + BM25 retrieval fused with RRF - three dsh tools (sema_search / sema_reindex / sema_stats) plus a sema CLI. It installs cleanly and boots without issues in our testing. It's early-stage but functional.

“Verified” means our automated CI actually ran dsh plugin add in a clean profile and it booted — nothing more. Feature descriptions and version compatibility are the author’s claims. This is not a security audit and not an endorsement of third-party code.

README

dsh-semantic-search

CI

Local semantic code search for DeepSeek Harness (dsh) — and any Node script.

中文文档:README.zh.md

sema builds a fragment-level index of a workspace: source is tokenized by a
language-aware tokenizer (camel/snake/kebab splitting, CJK n-grams), chunked with
symbol-aware boundaries (functions/classes stay intact), and embedded into fixed
dimension vectors — locally and dependency-free by default (feature-hashed
lexical TF-IDF), or via any OpenAI-compatible embedding endpoint. Queries run a
hybrid retrieval (vector cosine + BM25, fused with reciprocal-rank fusion) over
the index, so meaning-based search works even when exact terms don't match.

Ships as a dsh plugin bundle — the sema_search, sema_reindex and sema_stats
tools on the harness tool registry — plus a standalone sema CLI.


Highlights

  • Works offline by default — the built-in lexical provider needs no network,
    no model download, no API key. Use it as a fast BM25-plus code search.
  • Symbol-aware chunking — boundaries from a conservative per-language table
    (16 languages) keep functions/classes intact; a missed boundary degrades to a
    plain chunk rather than breaking.
  • CJK-aware tokenizer — n-gram tokenization (default bigrams) aligns Chinese
    queries and documents without a segmentation library; full-width punctuation is
    folded, not treated as a hard break.
  • Hybrid retrieval with RRF — vector cosine + BM25 channels are fused by
    reciprocal-rank fusion, so a document found by one channel still ranks.
  • Graceful degradation — if a configured remote provider is unreachable, the
    index falls back to the local lexical provider (configurable via allowFallback).
  • Incremental refresh + file watching — sema_reindex diffs by size+mtime,
    and an optional watcher keeps the index live.
  • Persistence — the index is saved to <root>/.sema atomically (JSON metadata
    • binary vectors), with staleness detection when the provider/dimension changes.
  • Deterministic — the same workspace and options produce the same index and
    the same ranked answers.

Supported languages

TypeScript, JavaScript, Python, Go, Rust, Java, Kotlin, Scala, C, C++, C#, Objective-C,
Ruby, PHP, Swift, Bash, plus common data/markup formats (JSON, YAML, TOML, Markdown,
HTML, XML...).


Installation

As a dsh bundle

The package declares "dsh": { "bundle": { "patch": "./cordis.patch.yml" } }. The
patch inserts one plugin row that mounts the bundle and registers
sema_search / sema_reindex / sema_stats on ctx.tools.

# from npm (name reserved; publish pending access setup)
npm install -g dsh-semantic-search

# straight from this repository
dsh plugin --profile demo add github:JohnXu22786/semantic-search

# or from a local checkout
dsh plugin --profile demo add /path/to/semantic-search

As a standalone CLI

npm install -g dsh-semantic-search   # or: npm run build && node bin/sema.mjs
sema --help

CLI usage

sema index                build the full index from the workspace
sema reindex [--full]     incremental refresh (or full rebuild with --full)
sema search <query...>    hybrid vector + BM25 search, prints top hits
sema stats [--json]       index health, provider, and sizing numbers

Global options:

--root <dir>          workspace root (default: current directory)
--data-dir <dir>      index storage directory (default: <root>/.sema)
--provider <kind>     embedding provider: lexical | openai (default: lexical)
--dim <n>             embedding dimension (lexical default: 4096; openai 0 = auto)
--base-url <url>      OpenAI-compatible embeddings endpoint base URL
--model <name>        embeddings model name (openai only)
--api-key <key>       API key (openai only; env: SEMA_EMBEDDING_API_KEY)
--top <n>             hits to print for search (default: 20)
--json                machine-readable output where supported
--help                show this help

CLI exit codes

  • 0 — success (including a search with zero hits and a --version/--help call).
  • 1 — a runtime failure (config error, build/index/search error).
  • 2 — a usage error: unknown command, unknown flag, or a missing query.

Configuration

The plugin is configurable through the bundle row's config (see
cordis.patch.yml for an example), the CLI flags above, or defaults in code:

Option Default Meaning
root cwd workspace root to index
dataDir .sema index storage directory
provider.kind lexical lexical (offline) or openai
provider.dimension 0 (lexical: 4096) embedding dimension; 0 = auto-infer from the endpoint
provider.baseUrl https://api.openai.com/v1 OpenAI-compatible endpoint root
provider.model text-embedding-3-small embedding model name
provider.apiKeyEnv SEMA_EMBEDDING_API_KEY env var holding the API key
allowFallback true fall back to the lexical provider when a remote one fails
include / ignore defaults glob sets of files to index / skip
maxLinesPerChunk 80 hard chunk size upper bound
nGram 2 CJK n-gram size (1 disables n-gramming)
topK 20 hits returned by default
rrfK 60 RRF fusion constant
vectorK 300 candidates per channel before fusion
autosave true persist the index after builds
autoIndex true build lazily on first search
watch true watch the workspace for changes

Development

npm ci
npm test          # build + run the node:test suite (76 tests)
npm run typecheck
npm run build     # tsc -> lib/

License

MIT — see LICENSE. © 2026 dsh-semantic-search contributors.

Install

🧩 Let your agent install it (recommended)

Install the catalog once, then DeepSeek Harness can find and install any plugin from this site automatically:

dsh plugin add dshbase-catalog

Then say "install semantic-search for me" — your agent finds it in the directory and installs it. Docs: dshbase-catalog · verified packs.

This plugin is GitHub source (not published to npm) — install it straight from the repo:

Web profile:

dsh plugin --profile web add github:JohnXu22786/semantic-search

Headless (CLI) profile:

dsh plugin --profile headless add github:JohnXu22786/semantic-search

Test report

Verified: L1 install + L2 load + L3 runtime from GitHub source on dsh 0.1.0-rc.6.

Security: not yet scanned — our daily static scan will cover it shortly.

Share this badge

More in Knowledge

Browse all 7819 plugins →