neo4j-vector-index-skill

When to Use

• Creating a vector index (CREATE VECTOR INDEX) on nodes or relationships
• Running vector similarity / nearest-neighbor search
• Storing embeddings on graph nodes during ingestion
• Indexing/querying embeddings already written by GDS algorithms
• Choosing similarity function, dimensions, HNSW params, or quantization
• Using SEARCH clause (2026.01+) or db.index.vector.queryNodes() (2025.x)
• Batch-updating embeddings after model change
• Combining vector results with immediate graph neighborhood (full retrieval_query pipelines → neo4j-graphrag-skill)
• Hybrid search that combines vector results with fulltext or other ranked sources

When NOT to Use

• GraphRAG pipelines (VectorCypherRetriever, HybridCypherRetriever, retrieval_query) → neo4j-graphrag-skill
• Fulltext-only / keyword-only search (FULLTEXT INDEX, db.index.fulltext.queryNodes) → neo4j-cypher-skill
• Computing GDS graph embeddings (FastRP, Node2Vec, GraphSAGE) → neo4j-gds-skill
• Index admin (list all indexes, drop range/text/lookup indexes) → neo4j-cypher-skill

---

Pre-flight — Determine Version

Drives syntax choice:

CALL dbms.components() YIELD versions RETURN versions[0] AS neo4j_version

Version	Use
2026.01 or higher	SEARCH clause (in-index filtering, preferred)
2025.x	db.index.vector.queryNodes() procedure (deprecated 2026.04 — use SEARCH when on 2026.x)

---

Step 1 — Create Vector Index

Node index (single label):

CYPHER 25
CREATE VECTOR INDEX chunk_embedding IF NOT EXISTS
FOR (c:Chunk) ON (c.embedding)
OPTIONS {
  indexConfig: {
    `vector.dimensions`: 1536,
    `vector.similarity_function`: 'cosine',
    `vector.quantization.type`: 'SCALAR',
    `vector.hnsw.m`: 16,
    `vector.hnsw.ef_construction`: 100
  }
}

Node index with filterable properties [2026.01+] — WITH declares which properties can be used in SEARCH ... WHERE:

CYPHER 25
CREATE VECTOR INDEX chunk_embedding IF NOT EXISTS
FOR (c:Chunk) ON (c.embedding)
WITH [c.source, c.lang, c.published_year]  // stored as metadata; filterable in SEARCH WHERE
OPTIONS { indexConfig: { `vector.dimensions`: 1536, `vector.similarity_function`: 'cosine' } }

Multi-label index with filterable properties [2026.01+]:

CYPHER 25
CREATE VECTOR INDEX doc_embedding IF NOT EXISTS
FOR (n:Document|Article) ON n.embedding
WITH [n.author, n.published_year, n.lang]
OPTIONS { indexConfig: { `vector.dimensions`: 1536, `vector.similarity_function`: 'cosine' } }

Relationship index:

CYPHER 25
CREATE VECTOR INDEX rel_embedding IF NOT EXISTS
FOR ()-[r:HAS_CHUNK]-() ON (r.embedding)
OPTIONS { indexConfig: { `vector.dimensions`: 768, `vector.similarity_function`: 'cosine' } }

WITH property types — only scalar types allowed: INTEGER, FLOAT, STRING, BOOLEAN, DATE, ZONED DATETIME, LOCAL DATETIME, ZONED TIME, LOCAL TIME, DURATION. Not allowed: LIST, POINT, or the vector property itself.

Index config reference:

Parameter	Type	Default	Notes
vector.dimensions	INTEGER 1–4096	none	Required; must match embedding model exactly
vector.similarity_function	STRING	'cosine'	'cosine' or 'euclidean'
vector.quantization.type	STRING	'binary' ('scalar' before 2026.08)	'none', 'scalar', 'binary' [2026.06+, GA 2026.07]; reduces storage; binary smallest (1 bit per dimension), most aggressive; needs vector-2.0+ (5.18+); set explicitly for version-independent behavior
vector.quantization.enabled	BOOLEAN	true	Deprecated 2026.06 — use vector.quantization.type; false without vector.quantization.type: 'none' fails index creation before 2026.07
vector.default_search_expansion_factor	FLOAT 1.0–10000.0	1.0 none / 1.5 scalar / 3.0 binary (was 2.0 before 2026.07)	[2026.06+, GA 2026.07]; value >1.0 on quantized vectors enables automatic rescoring with full-precision vectors (High-Fidelity Quantized search, HFQ); not settable at query time; existing indexes keep their stored value until rebuilt
vector.hnsw.m	INTEGER 1–512	16	HNSW graph connections; higher = better recall, more memory
vector.hnsw.ef_construction	INTEGER 1–3200	100	Build-time candidates; higher = better recall, slower build

Unquantized vectors: set vector.quantization.type: 'none' alone — on 2026.06, vector.quantization.enabled: false without it errors (fixed 2026.07).

Similarity function choice:

Use case	Function
Normalized embeddings (OpenAI, Cohere, Voyage, Google)	'cosine'
Unnormalized / raw distance matters	'euclidean'

Index providers — latest selected automatically; not specifiable in Cypher 25. Check with SHOW VECTOR INDEXES YIELD name, indexProvider:

Provider	Quantization support
vector-2026.08	High-Fidelity Quantized search for binary; 'binary' default
vector-2026.07	High-Fidelity Quantized search for scalar and binary
vector-2026.06	scalar and binary; required for binary + rescoring
vector-3.0 (2025.09+) / vector-2.0 (5.18+)	scalar

Changing quantization type, expansion factor, or provider requires drop + re-create + re-population. Indexes built before 2026.08 and rarely updated: re-create for 2026.09 rescored-binary/scalar speedup (~5× latency and throughput).

Memory — vector index files live in OS filesystem cache, not page cache. Leave RAM for: HNSW graph ≈ 8 B × vectors × vector.hnsw.m plus vector values (full precision ≈ 4 B × dims × vectors; scalar ÷4; binary ÷32). Size page cache for vector properties only if queries return or re-rank them. Details → Vector index memory configuration.

---

Step 2 — Wait for Index ONLINE

Index builds asynchronously — do NOT query until ONLINE:

SHOW VECTOR INDEXES YIELD name, state, populationPercent
WHERE name = 'chunk_embedding'
RETURN name, state, populationPercent

Poll every 5s until state = 'ONLINE' and populationPercent = 100.0. If state = 'FAILED' → stop, check logs.

Shell poll (cypher-shell):

until cypher-shell -u neo4j -p "$NEO4J_PASSWORD" \
  "SHOW VECTOR INDEXES YIELD name, state WHERE name='chunk_embedding' RETURN state" \
  | grep -q ONLINE; do
  sleep 5
done

---

Step 3 — Ingest Embeddings

Batch UNWIND pattern (use for > 100 nodes — never one-node-per-transaction):

from neo4j import GraphDatabase

driver = GraphDatabase.driver(uri, auth=(user, password))

def embed_batch(texts: list[str]) -> list[list[float]]:
    response = openai_client.embeddings.create(
        model="text-embedding-3-small", input=texts
    )
    return [r.embedding for r in response.data]

def store_embeddings(records: list[dict], batch_size: int = 500):
    expected_dim = 1536  # must match vector.dimensions
    texts = [r["text"] for r in records]
    embeddings = embed_batch(texts)
    for emb in embeddings:
        assert len(emb) == expected_dim, f"Dim mismatch: {len(emb)} != {expected_dim}"
    rows = [{"id": r["id"], "embedding": emb}
            for r, emb in zip(records, embeddings)]
    for i in range(0, len(rows), batch_size):
        driver.execute_query(
            "UNWIND $rows AS row MATCH (c:Chunk {id: row.id}) SET c.embedding = row.embedding",
            rows=rows[i:i+batch_size]
        )

❌ Never create index after embeddings are already stored — always create index first. ✅ Create index → poll ONLINE → ingest embeddings.

---

Step 4 — Run Vector Search

SEARCH clause (2026.01+, preferred)

CYPHER 25
MATCH (c:Chunk)
  SEARCH c IN (
    VECTOR INDEX chunk_embedding
    FOR $queryEmbedding
    LIMIT 10
  ) SCORE AS score
RETURN c.text, score
ORDER BY score DESC

With in-index filter [2026.01+] — properties must be declared in WITH at index creation:

// Index must have been created with: WITH [c.source, c.lang, c.published_year]
CYPHER 25
MATCH (c:Chunk)
  SEARCH c IN (
    VECTOR INDEX chunk_embedding
    FOR $queryEmbedding
    WHERE c.source = $source AND c.lang = 'en' AND c.published_year >= 2024
    LIMIT 10
  ) SCORE AS score
RETURN c.text, c.source, score
ORDER BY score DESC

Filtering strategy — choose one:

Strategy	When to use	Tradeoff
In-index WHERE [2026.01+]	Filters on pre-declared WITH properties; known at index design time	Fast, consistent latency; properties must be declared upfront
Post-filter (MATCH + procedure)	Arbitrary Cypher predicates, graph traversal, OR/NOT	Full flexibility; may over-fetch then discard
Pre-filter (MATCH first, then SEARCH)	Small known candidate set; exact nearest-neighbor within subset	Deterministic; slow on large candidate sets

In-index WHERE hard limits [2026.01+]:

• Property must be listed in WITH [...] at index creation — undeclared properties silently fall back to post-filtering
• AND predicates only — no OR, NOT, string ops. IN list membership allowed [2026.06+]
• Scalar types only: INTEGER, FLOAT, STRING, BOOLEAN, temporal types — not VECTOR/LIST/POINT

Post-filter pattern (2025.x or arbitrary predicates)

CYPHER 25
CALL db.index.vector.queryNodes('chunk_embedding', 50, $queryEmbedding)
YIELD node AS c, score
WHERE c.source = $source    // post-filter: fetch more, then filter
RETURN c.text, score
ORDER BY score DESC LIMIT 10

Relationship index procedure:

CYPHER 25
CALL db.index.vector.queryRelationships('rel_embedding', 5, $queryEmbedding)
YIELD relationship AS r, score
RETURN r.text, score

SEARCH clause hard limits (all versions):

• Index name cannot be a parameter ($indexName not allowed — use literal string)
• Binding variable must come from the enclosing MATCH pattern
• Query vector cannot reference the binding variable

---

Step 5 — Combine with Graph Traversal (simple cases)

Vector search as entry point, then graph hop:

CYPHER 25
MATCH (c:Chunk)
  SEARCH c IN (
    VECTOR INDEX chunk_embedding
    FOR $queryEmbedding
    LIMIT 10
  ) SCORE AS score
MATCH (c)<-[:HAS_CHUNK]-(a:Article)
OPTIONAL MATCH (a)-[:MENTIONS]->(org:Organization)
RETURN c.text, a.title, score, collect(DISTINCT org.name) AS organizations
ORDER BY score DESC

For full retrieval_query pipelines, HybridCypherRetriever, or neo4j-graphrag library → delegate to neo4j-graphrag-skill.

---

Step 6 — Hybrid Search

Use hybrid search when one signal misses useful candidates: semantic vectors miss exact terms, lexical fulltext misses paraphrases, structural graph signals find topology not present in text. The common pattern is vector + fulltext, but the same approach works for several vector indexes, GDS-written embeddings, graph traversal scores, or any two+ ranked/scored sources. Load references/hybrid-search.md and apply its query shape.

Rules:

• Run each source independently; rank each by score DESC, stable_id ASC.
• Combine by rank, not raw scores; fulltext and vector scores are not comparable.
• Every UNION ALL branch returns same columns: matched node + contribution.
• Use sourceK > finalK; combine before final limiting.
• Sum contributions per node; order final rows by wrrf DESC, stable_id ASC.
• Add more sources with extra UNION ALL branches and new sourceWeights keys.

---

Embedding Provider Quick-Reference

Provider / Model	Dimensions	Similarity	Notes
OpenAI text-embedding-3-small	1536	cosine	Default; reducible to 256–1536 via dimensions= param
OpenAI text-embedding-3-large	3072	cosine	Reducible to 256–3072
OpenAI text-embedding-ada-002	1536	cosine	Legacy; prefer 3-small
Cohere embed-v3 (English)	1024	cosine	Use input_type='search_document' at ingest, 'search_query' at query
Voyage voyage-3-large	1024	cosine	High quality; needs voyage-ai package
Google text-embedding-004	768	cosine	Via Vertex AI
Ollama nomic-embed-text	768	cosine	Local dev/testing
Ollama mxbai-embed-large	1024	cosine	Local; production-quality

vector.dimensions must exactly match model output — no auto-truncation.

---

Vector Functions

Ad-hoc similarity (not for kNN search — use index for that):

MATCH (a:Chunk {id: $id1}), (b:Chunk {id: $id2})
RETURN vector.similarity.cosine(a.embedding, b.embedding) AS sim
// vector.similarity.euclidean(a, b) — same signature, 0–1 range

// vector_distance (2025.10+) — metrics: EUCLIDEAN, EUCLIDEAN_SQUARED, MANHATTAN, COSINE, DOT, HAMMING
// Returns distance (lower = more similar, inverse of similarity)
RETURN vector_distance(a.embedding, b.embedding, 'COSINE') AS dist

// vector_dimension_count (2025.10+)
RETURN vector_dimension_count(n.embedding) AS dims

// vector_norm (2025.20+) — metrics: EUCLIDEAN, MANHATTAN
RETURN vector_norm(n.embedding, 'EUCLIDEAN') AS norm

Convert LIST to typed VECTOR:

// vector(value, dimension, coordinateType)
// coordinateType: FLOAT64, FLOAT32, INTEGER8/16/32/64
WITH vector([1.0, 2.0, 3.0], 3, 'FLOAT32') AS v
RETURN vector_dimension_count(v)

---

Index Management

// Show all vector indexes with config
SHOW VECTOR INDEXES YIELD name, state, populationPercent,
  labelsOrTypes, properties, indexConfig
RETURN name, state, populationPercent, labelsOrTypes, properties, indexConfig;

// Drop (node data unchanged — only index structure removed)
DROP INDEX chunk_embedding IF EXISTS;

// No ALTER VECTOR INDEX — to change dimensions or similarity function:
// 1. DROP INDEX old_index IF EXISTS
// 2. CREATE VECTOR INDEX new_index ... with new OPTIONS
// 3. Re-generate all embeddings with new model
// 4. Poll until ONLINE

---

Common Errors

Error	Cause	Fix
IllegalArgumentException: Index dimension mismatch	Stored embedding dim ≠ vector.dimensions	Fix embed generation; drop + recreate index with correct dim
Search returns incomplete results	Index still POPULATING	Poll until state = 'ONLINE'
Unknown procedure db.index.vector.queryNodes	Neo4j < 5.11	No vector index support below 5.11; upgrade
SEARCH clause not available	Neo4j < 2026.01	Use queryNodes() procedure
OR/NOT not allowed in SEARCH WHERE	SEARCH in-index filter restriction	Move complex predicates to outer WHERE after SEARCH
Zero results from correct query	Wrong similarity function or all-zeros embedding	Verify with vector.similarity.cosine(); check embed call succeeded
Score always 1.0	All-zeros or identical vectors	Embedding generation failed; add dimension assertion before ingest
vector.quantization.enabled / .type option rejected	provider vector-1.0 (Neo4j < 5.18)	Omit quantization option or upgrade to 5.18+
BINARY quantization rejected	provider older than vector-2026.06	Upgrade to 2026.06+; SHOW VECTOR INDEXES YIELD name, indexProvider to check (vector-2026.07 adds High-Fidelity Quantized search)

---

Checklist

• vector.dimensions matches embedding model output exactly
• Vector index created before ingesting embeddings
• Similarity function chosen explicitly (cosine for normalized, euclidean for distance-based)
• Index polled to state = 'ONLINE' before first query
• Dimension validated on every embedding before ingest
• SEARCH clause on Neo4j >= 2026.01 (preferred); procedure fallback only on 2025.x (deprecated 2026.04)
• SEARCH WHERE uses AND-only predicates with scalar types
• Batch UNWIND pattern used for > 100 nodes
• If model changes: drop index → recreate with new dimensions → re-generate all embeddings

---

In-Cypher Embedding Generation — ai.text.embed() [2025.12]

// Syntax (requires CYPHER 25)
CYPHER 25
// ai.text.embed(resource :: STRING, provider :: STRING, configuration :: MAP) :: VECTOR

Provider strings are lowercase ('openai', 'vertexai', 'bedrock-titan', 'azure-openai'). Full provider config → neo4j-genai-plugin-skill.

Full query pattern — embed at query time, search immediately (procedure fallback for 2025.x):

CYPHER 25
WITH ai.text.embed(
    "What are good open source projects",
    "openai",
    { token: $openaiKey, model: 'text-embedding-3-small' }) AS userEmbedding
CALL db.index.vector.queryNodes('chunk_embedding', 6, userEmbedding)  // deprecated 2026.04
YIELD node AS c, score
RETURN c.text, score
ORDER BY score DESC

With SEARCH clause (2026.01+):

CYPHER 25
WITH ai.text.embed("my query", "openai", { token: $openaiKey, model: 'text-embedding-3-small' }) AS userEmbedding
MATCH (c:Chunk)
  SEARCH c IN (VECTOR INDEX chunk_embedding FOR userEmbedding LIMIT 6) SCORE AS score
RETURN c.text, score
ORDER BY score DESC

Never pass API key as string literal — use $openaiKey parameter (inject via driver params) or apoc.static.get().

Rule: Use same model at ingest time and query time — embeddings from different models are not comparable.

Deprecated (still works but do not use in new code):

• genai.vector.encode() [deprecated] → use ai.text.embed() [2025.12]
• genai.vector.encodeBatch() [deprecated] → use CALL ai.text.embedBatch() [2025.12]
• genai.vector.listEncodingProviders() [deprecated] → use CALL ai.text.embed.providers() [2025.12]

For full ai.text.* reference (completion, structured output, chat, tokenization) → neo4j-genai-plugin-skill.

---

Cypher-Based Embedding Ingestion — db.create.setNodeVectorProperty

Set vector property via Cypher (e.g. during LOAD CSV or MERGE pipeline):

LOAD CSV WITH HEADERS FROM 'https://example.com/data.csv' AS row
MERGE (q:Question {text: row.question})
WITH q, row
CALL db.create.setNodeVectorProperty(q, 'embedding', apoc.convert.fromJsonList(row.question_embedding))

apoc.convert.fromJsonList() converts "[0.1,0.2,...]" STRING to LIST<FLOAT>. Python-generated embeddings → UNWIND batch pattern (Step 3).

---

Similarity Function — Extended Guidance

Choose based on training loss function:

• Check embedding model docs — models trained with cosine loss → use 'cosine'
• Models trained with L2/Euclidean loss → use 'euclidean'
• When docs are silent: default to 'cosine' (all major hosted APIs use it)

Common pitfall — wrong similarity function:

❌ Created index with 'euclidean' but model outputs L2-normalized vectors
   → scores are mathematically correct but rankings differ from expected cosine order
   → no error thrown; wrong results silently returned
✅ Verify: run vector.similarity.cosine(a.embedding, b.embedding) manually on known
   similar pairs — score should be > 0.9 for near-duplicate text

Sanity check query after index creation:

MATCH (c:Chunk) WITH c LIMIT 2
WITH collect(c) AS nodes
RETURN vector.similarity.cosine(nodes[0].embedding, nodes[1].embedding) AS cosine_check,
       vector.similarity.euclidean(nodes[0].embedding, nodes[1].embedding) AS euclidean_check

If both return null → embeddings not set. If cosine returns 1.0 → identical vectors (embed call failed).

---

Gotchas — Extended

Gotcha	Detail	Fix
Index not ONLINE at ingest time	Inserting nodes before index exists is valid — index auto-populates. But querying during POPULATING returns partial results	Always poll state = 'ONLINE' before first query
Wrong dimensions — silent failure	Stored vector dim ≠ vector.dimensions → IllegalArgumentException at query time, not at ingest time	Assert len(emb) == expected_dim before every SET c.embedding
Different models at ingest vs query	No error; cosine scores ~0.3–0.5 for clearly similar text	Use same model string/version for both; store model name as node metadata
Missing model at query	ai.text.embed returns null silently if provider config wrong	Test encode call standalone; check CYPHER 25 RETURN ai.text.embed(...) before embedding into pipeline
Large single-transaction ingest	One transaction for 10k nodes → OOM or timeout	Use UNWIND $rows ... CALL IN TRANSACTIONS OF 500 ROWS or Python batch loop
Chunk overlap not set	Adjacent chunks with no overlap → context at boundaries lost → poor recall for cross-paragraph queries	Set chunk_overlap ≥ 10% of chunk_size

---

References

Load on demand:

• Hybrid search - combine semantic, lexical, structural, or other ranked sources with WRRF/RRF
• Vector index docs
• SEARCH clause docs
• Vector functions docs
• ai.text.embed() / GenAI plugin docs [2025.12] — replaces deprecated genai.vector.encode()
• db.create.setNodeVectorProperty docs
• Chunking strategy, batch embed+store, splitter patterns — see document import skill
• Vector search with filters — 2026.01 preview