Redis Search

Single source of guidance for Redis Search — the retrieval surface that spans lexical, numeric, geo, JSON-path, and vector queries. Vector fields are part of the same FT.CREATE machinery as TEXT/TAG/NUMERIC fields, and FT.HYBRID blends lexical and vector ranking in one command, so this skill covers them together.

When to apply

• Creating, modifying, or reviewing a Redis Search index (FT.CREATE, FT.ALTER).
• Writing or optimizing FT.SEARCH, FT.AGGREGATE, or FT.HYBRID queries.
• Picking between TEXT, TAG, NUMERIC, GEO, GEOSHAPE, VECTOR, or JSON-path fields.
• Defining a VECTOR field, choosing HNSW vs FLAT, tuning HNSW parameters.
• Building a retrieval-augmented generation (RAG) pipeline.
• Rolling out a new index schema without downtime.
• Troubleshooting empty results, slow queries, or tokenization issues with FT.EXPLAIN, FT.PROFILE, FT.INFO.

1. Pick the right command

Three query commands. Reach for the narrowest one that fits.

# FT.SEARCH — most common
FT.SEARCH idx:products "@category:{electronics} @price:[100 500]" LIMIT 0 20 RETURN 3 name price category

# FT.AGGREGATE — top categories by avg price
FT.AGGREGATE idx:products "*" GROUPBY 1 @category REDUCE AVG 1 @price AS avg_price SORTBY 2 @avg_price DESC

# FT.HYBRID (Redis ≥ 8.4) — lexical + vector fusion
FT.HYBRID idx:docs
  SEARCH "@title:transformers" SCORER BM25 YIELD_SCORE_AS lexscore
  VSIM embedding $vec KNN count 1 K 50 YIELD_SCORE_AS vecscore
  COMBINE RRF 2 CONSTANT 60
  PARAMS 2 vec "..."
  DIALECT 2

For Redis < 8.4 the lexical+vector blend is approximated with FT.SEARCH pre-filter + =>[KNN ...]. See references/command-selection.md and references/hybrid-search.md.

2. Schema basics — FT.CREATE

FT.CREATE indexes Hash or JSON documents matching a PREFIX. Always set PREFIX. Use DIALECT 2 (the default since Redis 8; required for vector queries).

FT.CREATE idx:products ON HASH PREFIX 1 product:
    SCHEMA
        name TEXT WEIGHT 2.0
        category TAG SORTABLE
        price NUMERIC SORTABLE
        location GEO
        embedding VECTOR HNSW 6
            TYPE FLOAT32
            DIM 1536
            DISTANCE_METRIC COSINE

Pick the narrowest field type that supports your access pattern:

The classic mistake is TEXT for a category or status field "because it's a string" — TAG is roughly 10× faster for exact-match filtering.

See references/index-creation.md, references/field-types.md, references/dialect.md, references/ft-create-options.md, references/json-indexing.md.

3. Common queries

Narrow with filters; return only what you need.

# Tag filter + numeric range, sorted by price
FT.SEARCH idx:products "@category:{electronics} @price:[100 500]"
    SORTBY price ASC
    LIMIT 0 20
    RETURN 3 name price category

# Text + tag filter
FT.SEARCH idx:products "wireless headphones @category:{audio}"

# Negation and OR
FT.SEARCH idx:products "@category:{audio} -@brand:{generic} (@price:[0 100] | @on_sale:{true})"

Operators worth remembering: space = AND, | = OR, - = NOT, ~ = optional (scoring boost), =>{$weight: N} = boost. Escape hyphens and special characters inside TAG values (@sku:{ABC\\-123}). See references/query-syntax.md and references/search-syntax-primitives.md for the DSL vocabulary.

For tokenization gotchas (stemming, stopwords, language) see references/text-tokenization.md. For result shaping (SORTBY, RETURN, HIGHLIGHT, SUMMARIZE, NOCONTENT) see references/result-shaping.md. For performance levers (pre-filters, SORTABLE fields, tight RETURN, FT.PROFILE) see references/query-optimization.md.

4. Vector basics

Three vector settings have to match the embedding model exactly:

• DIM — output dimensionality (e.g. 1536 for OpenAI text-embedding-3-small). Mismatch produces silent garbage.
• DISTANCE_METRIC — COSINE for normalized text embeddings (common case), IP for unnormalized inner-product, L2 for raw Euclidean.
• TYPE — usually FLOAT32. Use FLOAT16 or quantized variants only when memory is the binding constraint.

# Index
FT.CREATE idx:docs ON HASH PREFIX 1 doc:
    SCHEMA
        content TEXT
        embedding VECTOR HNSW 6 TYPE FLOAT32 DIM 1536 DISTANCE_METRIC COSINE

# Pure KNN query (top 5 by cosine similarity)
FT.SEARCH idx:docs "*=>[KNN 5 @embedding $vec AS score]"
    PARAMS 2 vec "..."
    SORTBY score
    DIALECT 2

HNSW tuning levers: M (16–64, connections per node), EF_CONSTRUCTION (100–500, build quality), EF_RUNTIME (query-time candidate list).

See references/vector-query.md, references/algorithm-choice.md.

5. Hybrid retrieval

Two distinct patterns get called "hybrid." Pick by intent.

Filter-then-vector (any Redis version) — apply attribute filters so the engine narrows the search space before the vector comparison.

FT.SEARCH idx:docs "(@category:{tech} @date:[2024 +inf])=>[KNN 10 @embedding $vec AS score]"
    PARAMS 2 vec "..."
    SORTBY score
    DIALECT 2

Lexical + vector fusion (Redis ≥ 8.4) — blend BM25 text scoring with vector similarity, fuse with RRF or LINEAR. Use FT.HYBRID (see §1).

Don't fetch a wide unfiltered result and filter client-side — slower and less accurate. See references/hybrid-search.md.

6. Aggregations and shaping

FT.AGGREGATE is the declarative result-shaping command. Build a pipeline of stages.

# Top 5 categories by total revenue
FT.AGGREGATE idx:orders "@status:{shipped}"
    LOAD 2 @category @amount
    GROUPBY 1 @category
        REDUCE SUM 1 @amount AS revenue
    SORTBY 2 @revenue DESC
    LIMIT 0 5

Common stages: LOAD, APPLY (computed fields), FILTER (post-query), GROUPBY + REDUCE (SUM, COUNT, AVG, FIRST_VALUE, TOLIST), SORTBY, LIMIT.

For long-running result sets use WITHCURSOR + FT.CURSOR READ to page server-side. See references/aggregate-pipeline.md and references/aggregate-cursors.md.

7. RAG pattern

Standard pipeline: embed the query, vector-search Redis, pass top-K context to the LLM.

Practical tips:

• Match the metric to the embedding model (almost always COSINE for normalized text models).
• Chunk long documents (200–500-token chunks usually beat indexing whole pages).
• Batch inserts rather than one call per record.
• Pre-filter with attributes (tenant, recency, document type) before the vector search — see §5.
• Re-rank at the top of the funnel if precision matters more than recall.

See references/rag-pattern.md.

8. Operations

Zero-downtime schema changes: keep app queries pointed at an alias and swap the underlying index.

FT.CREATE idx:products_v2 ON HASH PREFIX 1 product: SCHEMA ...
FT.ALIASUPDATE products idx:products_v2
# App queries are stable:
FT.SEARCH products "@category:{electronics}"

Useful management commands: FT.INFO, FT.DROPINDEX, FT._LIST, FT.ALIASADD/UPDATE/DEL. See references/index-management.md.

Debug empty or slow queries with FT.EXPLAIN (shows how the query was parsed) and FT.PROFILE (shows execution stats). See references/debugging.md.

9. Client examples

Inline examples in this SKILL.md are CLI / RESP form — the wire protocol every client serializes to. For idiomatic snippets in a specific client:

• redis-py (Python, raw client): references/clients/python-redis-py.md
• Jedis (Java): references/clients/java-jedis.md
• RedisVL (Python, higher-level SDK on top of redis-py): references/clients/python-redisvl.md

Other clients (Lettuce, node-redis, go-redis, NRedisStack, .NET) translate the same CLI form; coverage is tracked as a follow-up.

References

• Redis: Search and query
• Redis: Vectors
• Redis: Query syntax
• Redis: Query dialects
• Redis: RAG quickstart
• FT.CREATE · FT.SEARCH · FT.AGGREGATE · FT.HYBRID
• RedisVL documentation