# Run retrieval on the machine, in-process or air-gapped

> Run retrieval on the machine, in-process or air-gapped: 2 companies fully do it, 10 with a limit, 17 carry the tag. Every listing in Retrieval & vector search with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/vector-search.local-embedded/
- Markdown: https://www.anchorterminal.com/niches/vector-search.local-embedded/index.md (~1,250 tokens)
- Slim: https://www.anchorterminal.com/niches/vector-search.local-embedded/index.min.md (~330 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/vector-search.local-embedded/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Two or three companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 2 | 10 | 17 | 6 | 2 |

## What counts as done

The index runs in-process or from a local directory with the same API as the server, holds metadata and supports updates, and needs no account or network call.

Measured, for our own tests: a 1M-chunk index opens and answers a query in under 2 seconds on a laptop.

How agents ask for it: keep the index local; search my files offline; no external service

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **LanceDB (LanceDB Systems, Inc.)**: The open-source library runs in-process on a local directory or object storage with metadata filters, updates and deletes, and needs no account. An account prefers it because the index never moves rows.
- **Milvus and Zilliz Cloud API + MCP (Zilliz Inc.)**: Milvus Lite runs locally through the same MilvusClient API with insert, upsert, delete, metadata filters and hybrid search, and no account.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Chroma API + MCP (Chroma Inc.)**: Runs in-process (ephemeral or persistent) with metadata and updates and no account, but hybrid, GroupBy and the Search API are Cloud only, the use case's named blocker.
- **Elasticsearch (Elasticsearch B.V.)**: Self-managed Elasticsearch runs locally with the same API and no account, but as a server, not in-process or from a directory.
- **Epsilla Vector Database (Epsilla, Inc.)**: The open-source database runs locally from a Docker image with the same pyepsilla client and no account, but as a server, not in-process.
- **Meilisearch (Meili SAS)**: The MIT Community Edition runs locally from one binary with the same API and no account, but as a local server, not in-process.
- **OpenSearch (OpenSearch Software Foundation, a project of The Linux Foundation)**: Runs locally in Docker under Apache-2.0 with the same API and no account, but as a server, not in-process; the whole HNSW graph sits in memory per an account.
- **Pinecone API + MCP (Pinecone Systems Inc.)**: The open-source pg_pinecone extension runs Pinecone's engine inside a self-managed Postgres 17 or 18 with no Pinecone call, but through SQL, not the cloud API, and not in-process.
- **Qdrant API + MCP (Qdrant Solutions GmbH)**: Qdrant Edge runs in-process like SQLite with no account, and the MCP server takes QDRANT_LOCAL_PATH, but Edge has its own API pages and a separate comparison with the server, so feature parity is not shown.
- **Typesense API + MCP (Typesense, Inc.)**: Self-hosted Typesense runs locally with the same API and no account, but as a server process, not in-process or from a directory.
- **Vespa (Vespa.ai AS)**: Runs locally in a Docker container with the same API and no account, but as a server that needs an application package deployed, not in-process.
- **Weaviate API + MCP (Weaviate B.V.)**: Embedded Weaviate runs locally with the same client API but is labelled experimental, and local instances can't use Weaviate Embeddings, so a local vectorizer must be configured.

- not this job: Cloudflare AI Search (Cloudflare, Inc.): Managed Cloudflare service only; no local or embedded mode.
- not this job: Upstash Vector API + MCP (Upstash, Inc.): Hosted service only; no self-hosted or embedded build.

## What stops agents finishing

- local mode lacks functions of the server (filters, hybrid)
- telemetry on by default in a self-hosted build

## Around it

Before: embeddings.local-model. After: vector-search.rag-over-docs.

## Where this comes from

Weight 6 of 100 in Retrieval & vector search: 2 of 37 accounts (air-gapped Qdrant embedded, LanceDB); kept small because most accounts were hosted or self-hosted servers

As data: https://www.anchorterminal.com/api/v1/use-cases/vector-search.local-embedded.json, or the MCP tool get_use_case with id vector-search.local-embedded. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
