Milvus

View as Markdown

A RocketRide store node that persists embedded documents in a Milvus vector database and retrieves them by semantic or keyword similarity search.

What it does

Stores pre-embedded document chunks in a Milvus collection and answers questions against them using semantic (vector) or keyword search. Supports both self-hosted Milvus and Zilliz Cloud. Documents must be run through an embedding node before reaching this node: ingesting a chunk without an embedding raises an error.

Uses pymilvus (MilvusClient). The collection is created automatically on first document ingest if it does not exist, with the vector dimension taken from the incoming embeddings. Re-ingesting a document replaces it: all existing entities with the same objectId are deleted before the new chunks are upserted in batches (default: 50 chunks per batch).

Documents removed at the source are soft-deleted by default. A markDeleted call flips the isDeleted metadata flag rather than dropping the entities, and searches exclude soft-deleted documents unless the filter explicitly asks for them. markActive reverses the flag if a document comes back. A hard remove deletes the entities outright.

When the node configuration is saved, the engine probes the connection (5-second timeout) and validates the collection name: 1-255 characters, letters, digits, and underscores only, starting with a letter or underscore.

---

Configuration

Lanes

Lane in	Lane out	Description
documents	(none)	Ingest pre-embedded documents into the collection
questions	documents	Return matching documents
questions	answers	Return matching documents as an answer
questions	questions	Enrich the question with matching documents for downstream nodes

Fields

Field	Type	Description
profile	string	Default "cloud". Connect to...
provider	string	Default "milvus".

Advanced settings

These keys are read from the node config but have no dedicated UI field:

Key	Default	Description
similarity	COSINE	Metric type for the vector index. Accepted values: L2, IP, COSINE, JACCARD, HAMMING, BM25. Any other value raises an error at startup.
timeout	60	Connection and operation timeout in seconds (minimum 1).
bulkInsertBatchSize	50	Number of chunks per bulk upsert batch (minimum 1).
renderChunkSize	33554432	Number of chunks fetched per group when rendering a full document.

---

Profiles

Profile	mode	Default host	Default port	Connection
Milvus cloud server (default)	cloud	(your Zilliz endpoint)	443	https://<host> with apikey as token
Your own Milvus server	local	localhost	19530	http://<host>:<port>, no token required

---

Collection schema & indexing

The auto-created collection has four fields:

Field	Type	Notes
id	INT64, primary key	Generated per chunk from the UUID1 timestamp combined with 27 random bits
vector	FLOAT_VECTOR	Dimension taken from the first ingested embeddings
content	VARCHAR (max length 65535)	The chunk text
meta	JSON	Chunk metadata used for filtering (objectId, nodeId, parent, chunkId, permissionId, isDeleted, and others)

Indexes: an IVF_FLAT vector index (nlist: 1024) using the configured similarity metric, and an STL_SORT scalar index on id.

---

Search behavior

Semantic search requires the question to carry an embedding (bind the pipeline to an embedding node) and verifies the embedding model matches the collection. Non-zero result offsets are not supported and raise an error. The query limit is raised to 25 whenever the requested limit is 10 or lower.

Scoring: with the COSINE metric, the raw Milvus distance is rescaled to [0, 1] via (distance + 1) / 2, where 1 means most similar. For other metrics a sigmoid of distance / -100 is used. Results below the score threshold are discarded.

Keyword search runs a substring match (content like '%query%') combined with the same metadata filters as semantic search.

All filter values (object IDs, node IDs, parents, permissions, and the keyword query) are escaped before interpolation into Milvus filter expressions.

---

Rendering

Given an objectId, the node can rehydrate the complete document text by fetching its chunks in chunkId order, in groups of renderChunkSize, and streaming the joined text to the output callback. Rendering only applies to objects that were vectorized by this pipeline (objects without a vector batch ID fall through to the next driver).

The store counts vectors (chunks), not documents: the node-reported document count is the number of entities in the collection.

---

Schema

Field	Type	Description	Default
milvus.profile	string	Type of Milvus host Connect to...	"cloud"
milvus.provider	string		const: "milvus"
vector.cloud.host		Enter the server IP address e.g. ..zillizcloud.com
vector.cloud.port			443
vector.local.host			"localhost"
vector.local.port			19530

Dependencies

• environs
• marshmallow
• grpcio
• milvus-lite ; platform_system != "Windows"
• pandas
• protobuf
• ujson
• pymilvus
• numpy