메뉴
Build vector search and embedding pipelines with Neumann. Use when implementing similarity search, RAG retrieval, semantic caching, or working with HNSW indexes.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Integrate Neumann database using official client SDKs. Use when writing Rust, Python, or TypeScript code that connects to Neumann, executes queries, or handles results.
Deploy and operate Neumann in production. Use when setting up single-node or cluster deployments, configuring Docker Compose or Kubernetes, or planning production infrastructure.
Model and query graph data in Neumann. Use when creating nodes and edges, traversing relationships, finding paths, running graph algorithms, or using Cypher-style queries.
Migrate data to Neumann from other databases. Use when moving data from PostgreSQL, MySQL, MongoDB, Neo4j, Pinecone, Redis, or other systems to Neumann.
Write correct Neumann database queries. Use when writing NODE, EDGE, EMBED, SIMILAR, ENTITY, VAULT, CACHE, BLOB, CHECKPOINT, CHAIN, or CLUSTER commands, or any Neumann SQL query.
Design Neumann data models using relational tables, graph nodes/edges, and vector embeddings. Use when planning a database schema or deciding which Neumann engines to use for a feature.
| name | neumann-vector |
| description | Build vector search and embedding pipelines with Neumann. Use when implementing similarity search, RAG retrieval, semantic caching, or working with HNSW indexes. |
Neumann stores embeddings as key-vector pairs in a dedicated vector engine backed by an HNSW (Hierarchical Navigable Small World) index for fast approximate nearest neighbor search.
Key concepts:
COSINE (default), EUCLIDEAN, DOT_PRODUCT.All vectors within a query must have the same dimensionality. Mixed dimensions cause runtime errors, not parse errors.
Embeddings are stored in tensor_store's embedding slab, which uses sharded
locking -- concurrent writes to different keys do not block each other.
The standard workflow: store embeddings, build the index, then query.
EMBED STORE 'doc-1' [0.12, -0.34, 0.56, 0.78]
Keys must be quoted strings. Unquoted keys with hyphens or colons will fail to parse.
EMBED STORE 'doc-1' [0.12, -0.34, 0.56] IN my_collection
EMBED BATCH [('doc-1', [0.1, 0.2, 0.3]), ('doc-2', [0.4, 0.5, 0.6]), ('doc-3', [0.7, 0.8, 0.9])]
Always prefer EMBED BATCH over multiple EMBED STORE calls for bulk inserts.
EMBED BUILD INDEX
This must be called after bulk inserts and before similarity queries. Building the index is an O(n log n) operation. Do it once after loading, not after each individual insert.
EMBED GET 'doc-1'
Returns the vector associated with the key, or an error if not found.
EMBED DELETE 'doc-1'
SHOW EMBEDDINGS
SHOW EMBEDDINGS LIMIT 50
COUNT EMBEDDINGS
SHOW VECTOR INDEX
The SIMILAR command performs approximate nearest neighbor search. It returns
key-score pairs ranked by similarity.
SIMILAR [0.1, 0.2, 0.3] LIMIT 10
SIMILAR 'doc-1' LIMIT 5
Uses the vector already stored under that key as the query vector.
SIMILAR [0.1, 0.2, 0.3] LIMIT 10 METRIC COSINE
SIMILAR [0.1, 0.2, 0.3] LIMIT 10 METRIC EUCLIDEAN
SIMILAR [0.1, 0.2, 0.3] LIMIT 10 METRIC DOT_PRODUCT
Default is COSINE if omitted.
SIMILAR [0.1, 0.2, 0.3] LIMIT 10 IN my_collection
SIMILAR [0.1, 0.2, 0.3] LIMIT 10 WHERE category = 'science'
The WHERE clause filters results after the ANN search.
Find similar embeddings that are also connected to a specific graph node:
SIMILAR [0.1, 0.2, 0.3] LIMIT 10 CONNECTED TO 'user-42'
This intersects ANN results with the graph neighborhood of user-42.
SIMILAR <key-or-vector> [LIMIT n] [METRIC COSINE|EUCLIDEAN|DOT_PRODUCT] [IN collection] [CONNECTED TO 'node-id'] [WHERE expr]
Result type: Similar (list of key + score pairs).
Find neighbors ranked by vector similarity:
NEIGHBORS 'node-1' BOTH BY SIMILARITY [0.1, 0.2, 0.3] LIMIT 5
This returns graph neighbors of node-1, re-ranked by cosine similarity to the
given vector.
SIMILAR [0.1, 0.2, 0.3] LIMIT 10 CONNECTED TO 'hub-node'
Restricts ANN results to embeddings whose keys correspond to nodes connected
to hub-node in the graph.
CACHE SEMANTIC PUT 'query-text' 'response-text'
CACHE SEMANTIC GET 'similar-query-text'
Uses embedding similarity for cache lookups -- returns cached responses for semantically similar (not just exact-match) queries.
ENTITY CREATE 'product-1' { name: 'Widget', price: 9.99 } EMBEDDING [0.1, 0.2, 0.3]
ENTITY UPDATE 'product-1' { price: 12.99 } EMBEDDING [0.15, 0.25, 0.35]
Stores properties, graph node, and embedding in one atomic operation.
Build index after bulk inserts. Call EMBED BUILD INDEX once after loading
all embeddings, not after each insert. Index building is expensive and
redundant if more inserts follow immediately.
Use EMBED BATCH for bulk loading. A single EMBED BATCH with 1000 pairs
is faster than 1000 individual EMBED STORE calls due to reduced parsing
and locking overhead.
Choose the right metric for your data:
COSINE -- best for normalized embeddings (most LLM outputs). Measures
angle between vectors, invariant to magnitude.EUCLIDEAN -- best for spatial data where absolute distance matters.DOT_PRODUCT -- best when magnitude carries meaning (e.g., popularity-weighted
embeddings). Fastest to compute but sensitive to vector norms.Use collections for multi-tenant isolation. Rather than filtering after search, partition embeddings into collections. Each collection maintains its own index, so queries only search relevant vectors.
Keep vector dimensions consistent. All vectors in a collection should have the same dimensionality. Mixing dimensions causes runtime errors.
LIMIT is not optional for large datasets. Without LIMIT, SIMILAR returns all embeddings sorted by distance. Always specify LIMIT for production queries.
Forgetting to build the index:
EMBED STORE 'k1' [0.1, 0.2]
SIMILAR [0.1, 0.2] LIMIT 5 -- slow: falls back to brute-force scan
EMBED BUILD INDEX -- should come before SIMILAR
SIMILAR [0.1, 0.2] LIMIT 5 -- fast: uses HNSW index
Using SIMILAR TO (wrong keyword):
SIMILAR [0.1, 0.2] LIMIT 5 -- correct
SIMILAR TO [0.1, 0.2] LIMIT 5 -- WRONG: no TO keyword in SIMILAR
Using BY instead of METRIC:
SIMILAR [0.1, 0.2] LIMIT 5 METRIC COSINE -- correct
SIMILAR [0.1, 0.2] LIMIT 5 BY COSINE -- WRONG: use METRIC
Unquoted keys with special characters:
EMBED STORE 'doc:1' [0.1, 0.2] -- correct
EMBED STORE doc:1 [0.1, 0.2] -- WRONG: colon splits into 3 tokens
EMBED STORE 'my-key' [0.1, 0.2] -- correct
EMBED STORE my-key [0.1, 0.2] -- WRONG: hyphen causes parse error
Searching without LIMIT on large datasets:
SIMILAR [0.1, 0.2] LIMIT 10 -- correct: bounded result set
SIMILAR [0.1, 0.2] -- risky: returns all embeddings