# Milvus for DSH

Milvus for DSH lets a DSH Web agent inspect and search a Milvus deployment
from chat. It supports Local Milvus and Zilliz Cloud, exact entity lookup,
scalar queries, BM25 full-text search, and dense+BM25 hybrid retrieval.
Natural-language dense search uses a DSH-managed embedding provider; BM25 runs
entirely from the collection's Milvus Function schema.

Every Milvus operation exposed by this plugin is read-only. The plugin does not
create collections, insert data, change indexes, or delete anything.

## Requirements

- DSH Web `0.1.0-rc.7` or later
- Node.js 22.19 or later
- A Milvus HTTP(S) endpoint reachable from the DSH Web host
- Optional: an API key from one of the supported embedding providers for dense
  and hybrid search

## Install

Install the package into the DSH Web profile:

```bash
dsh plugin --profile web add @zilliz/dsh-milvus
dsh web
```

If `dsh` is not installed globally:

```bash
npx --yes @deepseek-ai/dsh@0.1.0-rc.7 plugin --profile web add @zilliz/dsh-milvus
npx --yes @deepseek-ai/dsh@0.1.0-rc.7 web
```

After installing or updating the plugin, restart DSH Web and refresh the
browser page.

## Set up the plugin

Open **Settings → Plugins → Milvus for DSH**. The setup follows the same order
as using Milvus: connect a deployment, choose a collection, then enable only
the search capabilities you need.

### Connect Milvus

Choose **Local Milvus Standalone** or **Zilliz Cloud**, then enter the endpoint
and optional database. Local Milvus normally uses
`http://127.0.0.1:19530` and database `default`. Zilliz Cloud requires its HTTPS
endpoint and token. For an authenticated local deployment, select **Add
optional authentication** and enter its token.

After saving, use **Test connection**. The card collapses the form into a
connection summary so the deployment details no longer compete with collection
setup.

The endpoint is resolved from the machine running DSH Web. When Milvus runs in
another container or on another host, use an address reachable from the DSH
Web host—not a loopback address inside the Milvus container.

The active connection is bound when a new chat starts. Changing it affects new
chats; it does not silently switch an existing chat to a different deployment.

### Choose a collection

The Collection selector is populated from the connected Milvus database. Pick
one collection and DSH inspects its fields, indexes, and Functions on the Host.
You do not type collection or schema field names in the normal setup path.

The card then reports four capabilities:

- **Scalar query** is ready after a successful schema inspection.
- **BM25 search** is ready when the collection has one valid Milvus BM25
  Function route. It does not need an external API key.
- **Semantic search** is ready after its embedding provider and discovered
  `FloatVector` field are mapped.
- **Hybrid search** becomes ready automatically when both BM25 and semantic
  search are ready.

### Enable semantic search when needed

This step is required only for natural-language dense and hybrid search. BM25
text search does not use an external embedding provider.

1. Select **Enable** on the Semantic search capability.
2. Choose a provider and model that match the vectors already stored in the
   collection.
3. Enter the provider API key.
4. Choose one `FloatVector` field discovered from the selected collection.
5. Select **Enable semantic search**.

If a provider is already configured, reuse it instead of entering the key
again. The field must contain document vectors created with that exact model
and vector space. A matching dimension by itself does not prove compatibility;
for example, `gemini-embedding-001` and `gemini-embedding-2` are not
interchangeable.

Supported models in the settings UI:

| Provider | Models | Supported output dimensions |
| --- | --- | --- |
| OpenAI | `text-embedding-3-small`, `text-embedding-3-large`, `text-embedding-ada-002` | 1–1536; 1–3072; fixed 1536 |
| Google Gemini | `gemini-embedding-2`, `gemini-embedding-001` | 128–3072 |
| Cohere | `embed-v4.0`, `embed-english-v3.0`, `embed-english-light-v3.0`, `embed-multilingual-v3.0`, `embed-multilingual-light-v3.0` | v4: 256/512/1024/1536; v3 full: 1024; v3 light: 384 |
| Voyage AI | `voyage-4`, `voyage-4-large`, `voyage-4-lite`, `voyage-code-4`, `voyage-3.5`, `voyage-3.5-lite`, `voyage-code-3`, `voyage-finance-2`, `voyage-law-2` | 4/3.5/code-3: 256/512/1024/2048; finance/law: 1024 |
| Mistral AI | `mistral-embed`, `codestral-embed` | 1024; 1–3072 |
| Jina AI | `jina-embeddings-v5-text-small`, `jina-embeddings-v5-text-nano`, `jina-embeddings-v5-omni-small`, `jina-embeddings-v5-omni-nano`, `jina-embeddings-v4`, `jina-embeddings-v3` | small/v3: 32/64/128/256/512/768/1024; nano: through 768; v4: 128/256/512/1024/2048 |
| Together AI | `intfloat/multilingual-e5-large-instruct` | 1024 |

The first model shown for each provider is the recommended default. Older
models appear only when the provider still serves them and they are useful for
querying an existing collection created in that model's vector space. The
catalog intentionally excludes deprecated models and models that require a
dedicated endpoint.

The form disables vector fields whose dimensions the selected model cannot
produce. Dimension compatibility is necessary but not sufficient: the stored
document vectors must have been generated with that exact provider, model,
task mode, and vector space.

Milvus tokens and embedding API keys are saved through write-only DSH
Credentials. Their values are never stored in plugin settings, returned to the
browser after saving, or added to chat history.

The chat tool accepts natural-language query text; it never asks the user or
agent to supply a list of floats. DSH generates the query vector on the host,
checks its dimension against the collection schema, and sends it directly to
Milvus.

### Use Advanced settings only when necessary

**Advanced settings** is collapsed by default. Open it only to remove a vector
mapping, select among multiple schema-proven BM25 routes, or change hybrid
ranking.

No collection policy is needed when the collection has one valid BM25 route
and the default RRF with `k=60` is suitable. When several routes exist, select
one of the routes discovered from the schema; the UI does not accept arbitrary
text or sparse field names. You can also choose another RRF `k` or configure
named semantic/BM25 weights. A rerank parameter supplied in an individual chat
request takes precedence over the saved default.

## Collection requirements for search

The plugin inspects each collection before searching it and reports whether
dense, BM25, and hybrid retrieval are ready.

Dense search requires a configured binding to a dimensional `FloatVector`
field. BM25 search requires all of these collection facts:

- one analyzer-enabled `VarChar` or `TEXT` input field;
- a Milvus BM25 Function mapping that text field to a `SparseFloatVector`
  output field; and
- a BM25 index on that sparse field.

When exactly one valid BM25 route exists, the plugin selects it automatically.
If a collection has several BM25 text fields, the user must identify the one
to search in the request or save an exact collection policy; the agent does not
guess. A plain `SparseFloatVector` field without a BM25 Function is not enough
because the plugin cannot infer which external sparse encoder created it.

Hybrid search is ready only when one dense binding and one BM25 route are both
ready. It never silently falls back to one route if the other route is missing
or fails.

## Use it from chat

Start a new chat after activating the desired Milvus profile. A useful first
sequence is:

1. “List my Milvus collections.”
2. “Describe the `documents` collection.”
3. “Get IDs 10 and 11 from `documents`, returning `id`, `title`, and `source`.”
4. “Query `documents` where `year >= 2025`, returning `id` and `title`.”
5. “Search `documents` for documents about vector indexing, returning `id`,
   `title`, and `source`.”
6. “Use BM25 to search `documents` for the exact phrase `HNSW efConstruction`,
   returning `id`, `title`, and `source`.”
7. “Run hybrid search for `how HNSW indexing works`, returning `id` and
   `title`.”
8. “Run hybrid search with dense weight 0.7 and BM25 weight 0.3.”

The agent should discover and describe a collection before using its fields.
When a collection, field, partition, or filter is ambiguous, it should ask
rather than guess.

## Available tools

| Tool | Purpose |
| --- | --- |
| `milvus_list_collections` | List collections visible to the chat's bound profile. |
| `milvus_describe_collection` | Show schema, indexes, load state, and dense/BM25/hybrid readiness or blockers. |
| `milvus_get` | Retrieve up to 50 entities by exact Int64 or VarChar primary key. |
| `milvus_query` | Run a bounded scalar query with optional filter and partitions. |
| `milvus_search` | Embed natural-language query text and run bounded dense search with optional filter and partitions. |
| `milvus_text_search` | Run bounded natural-language BM25 search through a schema-proven Milvus BM25 Function. |
| `milvus_hybrid_search` | Combine configured dense and BM25 routes, then fuse their rankings with RRF or Weighted rerank. |

Data-retrieval tools return only requested scalar fields. Stored dense/sparse
vectors and generated query vectors are never returned to chat. The default
result limit is 10 and the maximum is 50.

Dense-search results include the Milvus distance, vector field and metric, the
embedding provider/model/dimension, and safe timing metadata. They do not
include the API key, raw provider error body, or generated vector.

Hybrid rerank is part of `milvus_hybrid_search`, not a separate tool:

- no rerank parameter: RRF with `k=60`;
- explicit RRF: the user may provide another positive `k`;
- explicit Weighted: the user must provide both `denseWeight` and
  `bm25Weight`, each from 0 to 1 and not both zero.

Named weights prevent route-order mistakes. If the user asks only for
“Weighted” without values, the agent asks for both weights instead of guessing.
Search results state the effective rerank values and whether they came from the
request, a collection policy, or the plugin default.

## What happens without an embedding key

Collection listing, description, exact get, scalar query, and schema-compatible
BM25 search continue to work. Dense and hybrid search are blocked, with a
specific configuration result:

- no collection binding: `retrieval_binding_absent`;
- binding refers to a missing provider profile: `embedding_profile_absent`;
- API key is missing or unavailable: `embedding_credential_unavailable`.

The plugin does not fall back to the chat model, another provider, a guessed
vector, or a scalar query.

## Privacy and safety

- Dense-search query text is sent from the DSH host to the embedding provider
  selected in the binding.
- The generated vector remains in host memory and is sent only to Milvus.
- Milvus tokens and provider keys stay behind the DSH Credentials boundary.
- Output fields must exist in the inspected schema and must be scalar.
- Filters may reference only scalar fields discovered from that collection.
- Exact route fields prevent a saved BM25 plan from silently switching routes.
  A saved schema fingerprint, when present, additionally blocks the plan after
  any retrieval-schema change until it is reviewed.
- The plugin exposes no mutation, schema, index, database, user, role, or
  administrative operation.
- External sparse encoders, model/cross-encoder rerank, ingestion, custom
  embedding endpoints, and manual vector input are not currently supported.

## Troubleshooting

### The settings card is missing

Confirm that the package is installed in the `web` profile, restart DSH Web,
and refresh the page:

```bash
dsh plugin --profile web why @zilliz/dsh-milvus
```

### A tool says no Milvus profile is available

Create a Milvus profile, make it active for new chats, and start a new chat.
Existing chats retain their original session binding.

### The Milvus connection test fails

Check host-to-Milvus network reachability, endpoint protocol and port, database
name, and token permissions. Local Milvus normally exposes HTTP on port `19530`.

### The embedding provider test fails

Check that the API key is configured and allowed to use the selected model.
Also check provider rate limits and outbound network access from the DSH Web
host.

### Dense search reports a dimension mismatch

Describe the collection and compare the bound field's dimension with the model
used during ingestion. Correct the binding or re-ingest with the intended
model; do not choose a different model only because it can produce the same
dimension.

### BM25 or hybrid search is blocked

Describe the collection and read its retrieval capability section. Common
blockers are `bm25_route_absent`, `bm25_route_ambiguous`,
`sparse_encoder_binding_absent`, `retrieval_plan_stale`, and
`retrieval_binding_absent`. Fix or re-save the collection policy,
Function/index, or dense binding; hybrid search does not degrade to a single
route.

## Update or remove

Update the package, restart DSH Web, and refresh the browser:

```bash
dsh plugin --profile web update @zilliz/dsh-milvus
dsh web
```

Remove it from the Web profile with:

```bash
dsh plugin --profile web remove @zilliz/dsh-milvus
```

Removing the plugin does not change or delete Milvus data. Review stored DSH
settings and credential records separately if they are no longer needed.

## Development

Install dependencies and run the local checks:

```bash
npm ci
npm test
npm pack --dry-run
```

Load a source checkout into DSH Web from this repository and restart DSH Web:

```bash
dsh plugin --profile web add "$PWD"
```

Read-only integration probes run only when an endpoint is supplied:

```bash
MILVUS_TEST_ENDPOINT=http://127.0.0.1:19530 npm run test:integration
MILVUS_TEST_ENDPOINT=http://127.0.0.1:19530 npm run test:integration:connection
```

Provider API smoke tests are separately network-gated. Set any supported key
in the environment, then opt in explicitly; providers without a key are
skipped, and neither keys nor returned vectors are printed:

```bash
EMBEDDING_TEST_ALLOW_NETWORK=1 npm run test:integration:embeddings
```

The mutation integration test creates, searches, and removes a disposable
fixture. Run it only against a non-production deployment after explicit opt-in:

```bash
MILVUS_TEST_ENDPOINT=http://127.0.0.1:19530 \
MILVUS_TEST_ALLOW_MUTATION=1 \
npm run test:integration:mutation
```

To verify the complete provider-to-Milvus path, also provide a Gemini API key.
This test embeds a query, searches a disposable 128-dimensional collection
through `milvus_search`, and removes the fixture:

```bash
MILVUS_TEST_ENDPOINT=http://127.0.0.1:19530 \
MILVUS_TEST_ALLOW_MUTATION=1 \
GEMINI_API_KEY=... \
npm run test:integration:retrieval
```

To verify BM25 and both hybrid rerank modes, use a non-production deployment.
The test first searches the existing `mfs_scale_2000` BM25 collection by
default, then creates and removes one disposable hybrid collection. Override
the existing collection name with `MILVUS_TEST_BM25_COLLECTION` when needed:

```bash
MILVUS_TEST_ENDPOINT=http://127.0.0.1:19530 \
MILVUS_TEST_ALLOW_MUTATION=1 \
npm run test:integration:hybrid
```

## License

Apache-2.0. See [LICENSE](./LICENSE).