Redis Search

RedisVL — Redis Search quick reference

This reference covers the search / index / query surface of RedisVL — the higher-level, schema-first Python SDK that builds on redis-py. It shows how RedisVL expresses the canonical CLI form via IndexSchema, SearchIndex, the query classes, and the FilterExpression DSL. It does not re-explain the query DSL grammar — that lives in ../search-syntax-primitives.md.

• redis-py (raw) equivalents for the same operations: python-redis-py.md. RedisVL is built on redis-py; an FT.* command that RedisVL does not wrap is reachable by calling redis-py directly through the index's underlying client (see §3). For tasks that are not RedisVL-specific (no schema, no vectorizer, no LLM primitive), prefer the redis-py reference.

• Query DSL vocabulary (delimiters, operators, escaping): ../search-syntax-primitives.md. RedisVL's FilterExpression is a typed builder for that DSL — the grammar itself is not duplicated here.

• Jedis (Java) equivalents for cross-language verification of FT.* behaviour: java-jedis.md. RedisVL is Python-only; there is no Java equivalent.

Examples below trace to the upstream RedisVL user-guide notebooks under redis/redis-vl-python/docs/user_guide/. Each section cites the originating notebook. The shared Bicycle dataset is reused where it fits (filters, ranges, RAG-shape examples). For vectorizer demos that need text-embedding-friendly prose, the upstream notebook's small sentence/document dataset is preserved verbatim.

Table of contents

1. When to choose RedisVL over raw redis-py

2. Minimum supported versions

3. Connection

4. Schema definition — Python dict

5. Schema definition — YAML

6. Storage type: HASH vs JSON

7. Index lifecycle

8. FilterExpression DSL

9. Query classes

10. Vectorizers

11. Hybrid retrieval

12. Async

13. LLM primitives (summary level)

14. Common errors & version gotchas

15. Upstream examples index

1. When to choose RedisVL over raw redis-py

redis-py equivalent: not applicable — see python-redis-py.md when none of the criteria below apply.

RedisVL builds on redis-py. The two are not mutually exclusive: when an FT.* command isn't wrapped at the RedisVL level, the same SearchIndex exposes its raw client via the public index.client property so an agent can fall through to the redis-py patterns described in python-redis-py.md without constructing a second connection. (The index._redis_client attribute also returns a client but is the internal lazy-init accessor — use .client in user code.)

2. Minimum supported versions

redis-py equivalent: see python-redis-py.md#1-minimum-supported-versions.

DIALECT default: RedisVL passes DIALECT 2 on every query by default — every query class accepts a dialect= parameter that defaults to 2 (VectorQuery, FilterQuery, RangeQuery, VectorRangeQuery, CountQuery, TextQuery). You do not need to set it explicitly. GEOSHAPE WITHIN/CONTAINS predicates still require manually overriding to dialect=3.

3. Connection

redis-py equivalent: see python-redis-py.md#2-connection-setup.

Three connection patterns, in order of preference. All three are equivalent in behaviour — pick the one that fits where the connection lives in your application.

from redisvl.index import SearchIndex
from redis import Redis

# (a) Pass a redis-py client. Use this when the connection is managed elsewhere
#     (DI container, app factory, test fixture). Mirrors 01_getting_started.ipynb cell 9.
client = Redis.from_url("redis://localhost:6379")
index = SearchIndex.from_dict(schema, redis_client=client, validate_on_load=True)

# (b) Pass a URL string. RedisVL constructs the client internally. Mirrors cell 11.
index = SearchIndex.from_dict(schema, redis_url="redis://localhost:6379")

# (c) Default — connects to redis://localhost:6379 when neither is set.
index = SearchIndex.from_dict(schema)

from_existing(name, ...) rehydrates a SearchIndex from a server-side index that already exists (the schema is read back via FT.INFO):

index = SearchIndex.from_existing("idx:bicycle", redis_url="redis://localhost:6379")

validate_on_load=True enables schema-shape validation on every .load() call — recommended in development, costly in tight write loops.

Falling through to raw redis-py. The underlying redis-py client is exposed for cases not covered by the high-level API:

raw = index.client                       # SyncRedisClient (a redis.Redis instance)
raw.ft("idx:bicycle").execute_command("FT.SOMETHING", ...)

When you reach for index.client, switch to python-redis-py.md for the idioms.

4. Schema definition — Python dict

redis-py equivalent: see python-redis-py.md#3-schema-imports and #5-create-index--json. RedisVL declares the schema once as data; raw redis-py passes a tuple of TextField / TagField / etc. constructors plus an IndexDefinition.

The canonical schema shape — mirrors 01_getting_started.ipynb cell 3:

schema = {
    "index": {
        "name": "user_simple",
        "prefix": "user_simple_docs",
        # "storage_type": "hash" (default) or "json" — see §6
    },
    "fields": [
        {"name": "user",         "type": "tag"},
        {"name": "credit_score", "type": "tag"},
        {"name": "job",          "type": "text"},
        {"name": "age",          "type": "numeric"},
        {
            "name": "user_embedding",
            "type": "vector",
            "attrs": {
                "dims": 3,
                "distance_metric": "cosine",
                "algorithm": "flat",
                "datatype": "float32",
            },
        },
    ],
}

The dict goes to SearchIndex.from_dict(schema) — see §3 for the connection variants.

Field types: tag, text, numeric, geo, vector. Each maps to the corresponding FT.CREATE schema clause.

Vector field algorithms: "flat" for exact / small-scale (≤ ~1M vectors), "hnsw" for ANN, "svs-vamana" for SVS Vamana (advanced; Intel-optimised; mention as optional — see 09_svs_vamana.ipynb).

The schema above uses 3-dim toy embeddings from the upstream 01_getting_started.ipynb notebook for brevity; the HNSW attrs block below shows the same schema shape with production-shape 1536 dims. dims must match whatever vectorizer or precomputed embedding produces the field's bytes — see §10.

HNSW attributes (added under attrs):

{
    "name": "embedding",
    "type": "vector",
    "attrs": {
        "dims": 1536,
        "distance_metric": "cosine",
        "algorithm": "hnsw",
        "datatype": "float32",
        "m": 16,                    # max bidirectional links per node
        "ef_construction": 200,     # build-time exploration factor
        "ef_runtime": 10,           # default query-time exploration factor
    },
},

JSON storage with nested paths. Set storage_type: "json" at the index level and use JSONPath-style nested field names (RedisVL emits the path-and-alias for you):

schema = {
    "index": {"name": "bike_index", "prefix": "bike:", "storage_type": "json"},
    "fields": [
        {"name": "name",            "type": "text"},
        {"name": "metadata.brand",  "type": "tag",     "path": "$.metadata.brand"},
        {"name": "metadata.price",  "type": "numeric", "path": "$.metadata.price"},
        # ...
    ],
}

The name is the query alias (@metadata.brand); the path is the underlying JSONPath. For top-level JSON fields the path defaults to $.<name>.

5. Schema definition — YAML

redis-py equivalent: not applicable. Raw redis-py has no YAML schema concept; you compose fields in code.

YAML is the recommended schema format for production — the file is checked into source control alongside the application and reused across sync/async/CLI consumers. Canonical shape — mirrors redis/redis-vl-python upstream docs/user_guide/schema.yaml:

version: '0.1.0'

index:
  name: vectorizers
  prefix: doc
  storage_type: hash

fields:
  - name: sentence
    type: text
  - name: embedding
    type: vector
    attrs:
      dims: 768
      algorithm: flat
      distance_metric: cosine

Load with from_yaml:

from redisvl.index import SearchIndex

index = SearchIndex.from_yaml(
    "schemas/bicycle.yaml",
    redis_url="redis://localhost:6379",
    validate_on_load=True,
)

Field shape under fields: mirrors the dict form in §4 — same type, attrs, path. The top-level version: field is the schema-format version, not the data version; upstream docs/user_guide/schema.yaml at tag v0.18.2 pins it to '0.1.0'. RedisVL validates this format-version string at parse time — keep it '0.1.0' for RedisVL 0.18.x and bump only when a future RedisVL release introduces a new schema-format major.

6. Storage type: HASH vs JSON

redis-py equivalent: see python-redis-py.md#4-create-index--hash and #5-create-index--json.

Mirrors 05_hash_vs_json.ipynb. Set storage_type once in the schema; everything else follows.

# HASH storage — default. Each Redis key holds a flat hash; field names are the schema names verbatim.
hash_schema = {
    "index": {"name": "user-hash", "prefix": "user-hash-docs", "storage_type": "hash"},
    "fields": [
        {"name": "user",            "type": "tag"},
        {"name": "office_location", "type": "geo"},
        {"name": "user_embedding",  "type": "vector",
         "attrs": {"dims": 3, "distance_metric": "cosine",
                   "algorithm": "flat", "datatype": "float32"}},
    ],
}
hindex = SearchIndex.from_dict(hash_schema, redis_url="redis://localhost:6379")
hindex.create(overwrite=True)
hindex.storage_type        # -> StorageType.HASH

# JSON storage — same fields, nested JSON document per key. Use JSONPath for nested fields.
json_schema = {
    "index": {"name": "user-json", "prefix": "user-json-docs", "storage_type": "json"},
    "fields": [
        {"name": "user",            "type": "tag"},
        {"name": "office_location", "type": "geo"},
        {"name": "user_embedding",  "type": "vector",
         "attrs": {"dims": 3, "distance_metric": "cosine",
                   "algorithm": "flat", "datatype": "float32"}},
    ],
}
jindex = SearchIndex.from_dict(json_schema, redis_url="redis://localhost:6379")
jindex.create(overwrite=True)

Choosing between them:

When in doubt for AI/RAG workloads, JSON: it preserves the natural shape of LLM outputs and allows nested metadata without flattening conventions. HASH wins when documents are already flat key-value structures.

7. Index lifecycle

redis-py equivalent: see python-redis-py.md#12-index-management. RedisVL wraps the same FT.CREATE / FT.DROPINDEX / FT.ALTER / FT.ALIAS* commands behind methods on SearchIndex.

# Create the index. overwrite=True drops and recreates; drop=True (with overwrite=True)
# also drops every existing indexed document.
index.create(overwrite=True, drop=False)

# Existence + introspection
index.exists()              # -> bool
index.info()                # -> dict from FT.INFO

# Load documents. Returns the list of full Redis keys written.
keys = index.load(data)                          # auto-generates document IDs
keys = index.load(data, id_field="user")         # use a field as the document ID
keys = index.load(data, ttl=3600)                # set per-document TTL

# Fetch by document ID (the unprefixed ID — RedisVL prepends the index prefix).
record = index.fetch("john")

# Delete by full key (with prefix) or by document ID (RedisVL adds the prefix).
index.drop_keys("user_simple_docs:01ABC...")
index.drop_documents("john")
index.drop_documents(["mary", "joe"])

# Clear: delete every indexed document but keep the index.
n_deleted = index.clear()

# Delete the index. drop=True (default) also deletes all indexed documents (FT.DROPINDEX ... DD);
# drop=False keeps the documents and removes only the index definition.
index.delete(drop=True)

index.load(data) is the bulk-write entry point. It accepts a list of dicts; each dict's keys must match the schema's field names (or the JSON path aliases for JSON-storage schemas). For HASH-storage vector fields the dict value must be np.ndarray(...).astype(np.float32).tobytes() (or use a vectorizer, see §10). For JSON-storage vector fields a plain list[float] works.

Updating existing documents. RedisVL 0.18.2 does not expose a dedicated update_load(...) method (the name appears in early proposals; it is not in the shipped API). To update existing indexed documents, call .load(...) again with explicit keys= matching the existing Redis keys, or with the same id_field value:

# Re-load the same document IDs. HASH storage merges field-by-field at the Redis
# level (HSET semantics); JSON storage replaces the document.
keys = index.load(updated_data, id_field="user")

# Or write to explicit keys (must be full prefixed keys).
keys = index.load(updated_data, keys=["user_simple_docs:01ABC", "user_simple_docs:01DEF"])

The preprocess= parameter accepts a callable applied to each item before write — useful for normalising shape or computing derived fields. validate_on_load=True (set on the index) re-runs schema validation on every .load() call.

paginate(query, page_size=N) is the recommended way to walk large result sets without holding everything in memory — yields batches of result dicts:

from redisvl.query import FilterQuery
from redisvl.query.filter import FilterExpression

query = FilterQuery(filter_expression=FilterExpression("*"), return_fields=["user", "age", "job"])
for batch in index.paginate(query, page_size=100):
    for doc in batch:
        process(doc)

Mirrors 01_getting_started.ipynb cells 13, 18, 23, 26, 28, 30, 32, 33.

8. FilterExpression DSL

redis-py equivalent: see python-redis-py.md#6-ftsearch-idioms for raw query strings. The full DSL grammar lives in ../search-syntax-primitives.md; this section shows only how RedisVL's typed FilterExpression compiles to that grammar.

The four filter classes mirror the four indexable scalar types — Tag, Text, Num, Geo — plus Timestamp (a Num subclass that accepts datetime objects). All five live in redisvl.query.filter. Operators are overloaded: ==, !=, <, >, <=, >=, % (text wildcard/fuzzy), & (AND), | (OR), ~ (NOT — used via !=).

Tag — mirrors 02_complex_filtering.ipynb cells 8, 10, 12

from redisvl.query.filter import Tag

Tag("credit_score") == "high"                       # @credit_score:{high}
Tag("credit_score") != "high"                       # -@credit_score:{high}
Tag("credit_score") == ["high", "medium"]           # @credit_score:{high|medium}
Tag("credit_score") == set(["high", "medium"])      # same; set enforces uniqueness
Tag("credit_score") == []                           # gracefully -> "*" (no constraint)

Num — mirrors cells 17, 18, 19

from redisvl.query.filter import Num

Num("age").between(15, 35)        # @age:[15 35]
Num("age") == 14                  # @age:[14 14]
Num("age") != 14                  # -@age:[14 14]
Num("age") < 18                   # @age:[-inf (18]
Num("age") >= 18                  # @age:[18 +inf]

Text — mirrors cells 25, 26, 27, 28, 29

from redisvl.query.filter import Text

Text("job") == "doctor"           # @job:"doctor"
Text("job") != "doctor"           # -@job:"doctor"
Text("job") % "doct*"             # @job:doct*  (wildcard / prefix)
Text("job") % "%%engine%%"        # @job:%%engine%%  (fuzzy, Levenshtein 2)
Text("job") % "engineer|doctor"   # @job:(engineer|doctor)
Text("job") % ""                  # gracefully -> "*"

Geo — mirrors cells 34, 35, 36

from redisvl.query.filter import Geo, GeoRadius

Geo("office_location") == GeoRadius(-122.4194, 37.7749, 10, "km")    # @office_location:[-122.4194 37.7749 10 km]
Geo("office_location") != GeoRadius(-122.4194, 37.7749, 10, "km")    # -@office_location:[...]

Timestamp — mirrors cells 21–23

from redisvl.query.filter import Timestamp
from datetime import datetime

dt = datetime(2025, 3, 16, 13, 45, 39)
Timestamp("last_updated") > dt                    # @last_updated:[(<epoch> +inf]
Timestamp("last_updated").between(dt1, dt2)       # @last_updated:[<epoch1> <epoch2>]

Timestamp converts datetime to epoch seconds and emits a NUMERIC range — the underlying field must be declared type: numeric and stored as epoch seconds.

Composition — mirrors cells 38, 40, 42

Boolean operators compose any filter into a single expression:

t  = Tag("credit_score") == "high"
lo = Num("age") >= 18
hi = Num("age") <= 100
ts = Timestamp("last_updated") > datetime(2025, 3, 16, 13, 45, 39)

combined = t & lo & hi & ts                       # AND
either   = (Num("age") < 18) | (Num("age") > 93)  # OR

# Defensive composition — empty filters fall back to "*", so partial inputs compose cleanly.
def make_filter(age=None, credit=None, job=None):
    return (
        (Num("age") > age) &
        (Tag("credit_score") == credit) &
        (Text("job") % job)
    )

Pass the resulting FilterExpression to any query class via filter_expression=:

from redisvl.query import VectorQuery

v = VectorQuery(
    vector=[0.1, 0.1, 0.5],
    vector_field_name="user_embedding",
    return_fields=["user", "credit_score", "age", "job", "office_location"],
    filter_expression=combined,
)

Or swap the filter on an existing query without rebuilding it:

v.set_filter(Tag("credit_score") != "high")

str(filter_expression) returns the compiled query string — useful for logging or for verifying what RedisVL emits before sending it to Redis.

9. Query classes

redis-py equivalent: see python-redis-py.md#6-ftsearch-idioms (FT.SEARCH) and #7-ftaggregate-idioms (FT.AGGREGATE). RedisVL's query classes route to either depending on shape.

All seven live under redisvl.query. Pass an instance to index.query(...) — RedisVL picks the right underlying command:

VectorQuery — KNN with optional pre-filter

Mirrors 01_getting_started.ipynb cell 36 + 02_complex_filtering.ipynb cell 8.

from redisvl.query import VectorQuery
from redisvl.query.filter import Tag, Num

# Top-3 nearest, no filter.
query = VectorQuery(
    vector=[0.1, 0.1, 0.5],
    vector_field_name="user_embedding",
    return_fields=["user", "age", "job", "credit_score", "vector_distance"],
    num_results=3,
)
results = index.query(query)

# Pre-filtered KNN — filter is applied before the vector scan.
query = VectorQuery(
    vector=[0.1, 0.1, 0.5],
    vector_field_name="user_embedding",
    return_fields=["user", "credit_score", "age", "vector_distance"],
    num_results=3,
    filter_expression=(Tag("credit_score") == "high") & (Num("age").between(18, 60)),
)
results = index.query(query)

results is a list[dict] — each dict carries the returned fields plus id and (when the vector_distance alias is in return_fields) the distance score.

VectorRangeQuery — within a distance threshold

from redisvl.query import VectorRangeQuery
# Backwards-compatible alias — RangeQuery is a thin subclass kept for older code paths.
from redisvl.query import RangeQuery  # `class RangeQuery(VectorRangeQuery): pass` in redisvl/query/query.py

range_q = VectorRangeQuery(
    vector=[0.1, 0.1, 0.5],
    vector_field_name="user_embedding",
    return_fields=["user", "vector_distance"],
    distance_threshold=0.5,        # COSINE distance in [0, 2]
    num_results=10,                 # cap on returned results
)
results = index.query(range_q)

Mirrors docs/user_guide/11_advanced_queries.ipynb (distance-threshold pattern; the notebook uses the equivalent filter_expression + VectorQuery shape rather than VectorRangeQuery directly — the class is exercised by tests/integration/test_query.py).

FilterQuery — pure filter retrieval

Mirrors 01_getting_started.ipynb cell 30.

from redisvl.query import FilterQuery
from redisvl.query.filter import Tag

query = FilterQuery(
    filter_expression=Tag("credit_score") == "high",
    return_fields=["user", "age", "job"],
    num_results=50,
)

# Paginate large result sets — recommended over num_results for full scans.
for batch in index.paginate(query, page_size=100):
    for doc in batch:
        process(doc)

CountQuery — count without fetching

from redisvl.query import CountQuery
from redisvl.query.filter import Tag

count = index.query(CountQuery(filter_expression=Tag("brand") == "Nike"))
# Returns the integer match count; emits FT.SEARCH ... LIMIT 0 0 under the hood.

Source: redisvl.query.CountQuery (redisvl/query/query.py — class at module scope). Exercised by tests/integration/test_query.py; the user-guide notebooks favour FilterQuery over CountQuery for didactic reasons (showing the fetched docs), but the class is a stable part of the v0.18.2 public API.

TextQuery — full-text scored retrieval

Mirrors 11_advanced_queries.ipynb cells 8, 10, 13.

from redisvl.query import TextQuery
from redisvl.query.filter import Tag, Num

# Plain text search with BM25STD (default scorer).
text_query = TextQuery(
    text="running shoes",
    text_field_name="brief_description",
    return_fields=["product_id", "brief_description", "category", "price"],
    num_results=5,
)

# Add a filter expression to scope the search.
filtered = TextQuery(
    text="comfortable",
    text_field_name="brief_description",
    filter_expression=Num("price") < 100,
    return_fields=["product_id", "brief_description", "price"],
)

# Multi-field weighting — search across multiple text fields with per-field weights.
weighted = TextQuery(
    text="shoes",
    text_field_name={"brief_description": 1.0, "full_description": 0.5},
    return_fields=["product_id", "brief_description"],
)

# Stopword handling — "english" is the default; pass a custom list or None to disable.
with_stopwords = TextQuery(text="the best shoes", text_field_name="brief_description", stopwords="english")
custom_stop    = TextQuery(text="best shoes", text_field_name="brief_description", stopwords=["for", "with"])
no_stopwords   = TextQuery(text="the best shoes", text_field_name="brief_description", stopwords=None)

Scorers: TFIDF, TFIDF.DOCNORM, BM25STD (default), BM25STD.NORM, BM25STD.TANH, DISMAX, DOCSCORE, HAMMING. See 11_advanced_queries.ipynb cells 10–11.

AggregationQuery — FT.AGGREGATE shape

AggregationQuery lives at redisvl.query.AggregationQuery (redisvl/query/aggregate.py) and subclasses redis-py's AggregateRequest. Its constructor takes a single query_string (the FT.AGGREGATE filter expression) and the inherited fluent API matches the one in python-redis-py.md#7-ftaggregate-idioms — .load(), .apply(), .group_by(), .filter(), .sort_by(), .limit(), .cursor().

from redisvl.query import AggregationQuery
from redis.commands.search import reducers

# Top-3 brands by document count, descending — `*` matches all docs.
agg = (
    AggregationQuery("*")
    .group_by("@brand", reducers.count().alias("n"))
    .sort_by(("@n", "DESC"))
    .limit(0, 3)
)
results = index.aggregate(agg)

Route through index.aggregate(...) when you want the RedisVL result-shape normalisation; the underlying command is FT.AGGREGATE on the index's name. The reducer factories live in redis.commands.search.reducers (imported from redis-py) — RedisVL does not re-export them. redisvl/query/aggregate.py declares only the subclass plus AggregateHybridQuery, MultiVectorQuery, and Vector; the fluent surface is inherited verbatim. The user-guide notebooks (11_advanced_queries.ipynb) cover the related AggregateHybridQuery (cells 34, 37, 42) but do not include a worked AggregationQuery example — see tests/integration/test_aggregation.py for executable references.

SVS Vamana (advanced / optional)

09_svs_vamana.ipynb covers Intel's SVS Vamana vector algorithm. Treat it as advanced/optional in v1: it requires a Redis build with SVS support (Intel-optimised) and adds tuning parameters orthogonal to the HNSW/FLAT trade-offs. Declare it via "algorithm": "svs-vamana" in the vector field's attrs. Most agents should pick FLAT (≤ ~1M vectors) or HNSW (ANN at scale) and only reach for SVS Vamana when explicitly asked.

10. Vectorizers

redis-py equivalent: not applicable. Raw redis-py has no vectorizer concept — you compute the embedding yourself and pass the raw bytes blob into the query (np.array(...).astype(np.float32).tobytes()). RedisVL's vectorizers wrap that pattern plus the provider-specific HTTP/SDK call.

Canonical example — OpenAI

Mirrors 04_vectorizers.ipynb cells 4–7:

import os
from redisvl.utils.vectorize import OpenAITextVectorizer

oai = OpenAITextVectorizer(
    model="text-embedding-ada-002",                   # or text-embedding-3-small / -3-large
    api_config={"api_key": os.environ["OPENAI_API_KEY"]},
)

# Single embedding.
vec = oai.embed("This is a test sentence.")
print(len(vec))                                       # 1536 for ada-002 / 3-small

# Batch.
sentences = ["That is a happy dog", "That is a happy person", "Today is a sunny day"]
embeddings = oai.embed_many(sentences)

# Async batch — vectorizers expose .aembed_many and .aembed for async use.
embeddings = await oai.aembed_many(sentences)

Vectorizers all expose the same surface: .embed(text), .embed_many(texts), .aembed(text), .aembed_many(texts). Pass as_buffer=True to embed_many to get RedisVL's binary buffer format suitable for direct write into HASH-storage vector fields without the np.array(...).tobytes() conversion.

Provider table

All providers live under redisvl.utils.vectorize. Mirrors redisvl/utils/vectorize/__init__.py.

RedisVL vectorizers auto-detect the model's dim on first call for known providers — you only need to set dims in the schema when the vectorizer doesn't expose it (e.g., custom). When dims in the schema disagrees with what the vectorizer produces, .load() raises a dimension-mismatch error at write time.

11. Hybrid retrieval

redis-py equivalent: see python-redis-py.md#10-fthybrid. RedisVL exposes the same FT.HYBRID surface through HybridQuery and AggregateHybridQuery.

Version gate: HybridQuery requires Redis ≥ 8.4.0 and redis-py >= 7.1.0. On older Redis or older redis-py, fall back to the pre-filter + KNN pattern via VectorQuery(filter_expression=...) — see below. The upstream 11_advanced_queries.ipynb gates every HybridQuery example on a HYBRID_SEARCH_AVAILABLE flag (cells 32, 36, 39, 41).

Native FT.HYBRID — HybridQuery

Mirrors 11_advanced_queries.ipynb cell 32:

from redisvl.query import HybridQuery

hybrid_query = HybridQuery(
    text="running shoes",
    text_field_name="brief_description",
    vector=[0.1, 0.2, 0.1],                       # query vector
    vector_field_name="text_embedding",
    return_fields=["product_id", "brief_description", "category", "price"],
    num_results=5,
    yield_text_score_as="text_score",
    yield_vsim_score_as="vector_similarity",
    combination_method="LINEAR",                  # or "RRF" (server default)
    linear_alpha=0.3,                             # 30% text, 70% vector
    yield_combined_score_as="hybrid_score",
)

results = index.query(hybrid_query)

Combination methods:

• combination_method="RRF" — Reciprocal Rank Fusion. Rank-based, robust without weight tuning. Knobs: rrf_window (default 20), rrf_constant (default 60).

• combination_method="LINEAR" — weighted score blend. Knob: linear_alpha (text weight; 1 - alpha is the implicit vector weight).

Scorers (text leg): same options as TextQuery — BM25STD (default), TFIDF, DISMAX, etc.

Filter expressions apply to both legs:

filtered_hybrid = HybridQuery(
    text="professional equipment",
    text_field_name="brief_description",
    vector=[0.9, 0.1, 0.05],
    vector_field_name="text_embedding",
    filter_expression=Num("price") > 100,
    combination_method="LINEAR",
    yield_text_score_as="text_score",
    yield_vsim_score_as="vector_similarity",
    yield_combined_score_as="hybrid_score",
)

Aggregate-shaped hybrid — AggregateHybridQuery

When you want the FT.AGGREGATE-shape output (rows, group-by stages) instead of the FT.SEARCH-shape:

from redisvl.query import AggregateHybridQuery

agg_hybrid = AggregateHybridQuery(
    text="running shoes",
    text_field_name="brief_description",
    vector=[0.1, 0.2, 0.1],
    vector_field_name="text_embedding",
    return_fields=["product_id", "brief_description", "category", "price"],
    alpha=0.7,                                    # 70% vector, 30% text
    num_results=5,
)
results = index.query(agg_hybrid)

Fallback for Redis < 8.4.0 — pre-filter + KNN

When FT.HYBRID is unavailable, blend retrieval by applying the text filter as a pre-filter on the vector query:

from redisvl.query import VectorQuery
from redisvl.query.filter import Text, Num

query = VectorQuery(
    vector=query_vector,
    vector_field_name="text_embedding",
    return_fields=["product_id", "brief_description", "price", "vector_distance"],
    num_results=10,
    # Pre-filter shrinks the candidate set before the KNN scan.
    filter_expression=(Text("brief_description") % "running") & (Num("price") < 200),
)
results = index.query(query)

This is not the same operation as FT.HYBRID: it filters by text but ranks purely by vector distance. Use it when (a) the Redis version doesn't support FT.HYBRID, or (b) the workload only needs filtered KNN, not blended ranking. The upstream 02_complex_filtering.ipynb cell 8 demonstrates the pattern in its native shape.

12. Async

redis-py equivalent: see python-redis-py.md footer (async out of scope for v1).

AsyncSearchIndex mirrors SearchIndex method-for-method on the search/index/query surface. Async parity is at full coverage for the operations relevant to this reference — create, delete, load, fetch, query, paginate (async generator), aggregate, search, clear, drop_keys, drop_documents, expire_keys, exists, info, listall.

from redisvl.index import AsyncSearchIndex
from redis.asyncio import Redis

client = Redis.from_url("redis://localhost:6379")
index = AsyncSearchIndex.from_dict(schema, redis_client=client)

await index.create(overwrite=True, drop=False)
await index.load(data)

# Same query classes; same construction.
from redisvl.query import VectorQuery
query = VectorQuery(vector=[0.1, 0.1, 0.5], vector_field_name="user_embedding",
                    return_fields=["user", "age"], num_results=3)
results = await index.query(query)

await index.delete(drop=True)

Mirrors 01_getting_started.ipynb cells 41–47.

Divergences from the sync API:

• from_existing(...) is async on AsyncSearchIndex — await AsyncSearchIndex.from_existing(name, redis_url=...). Sync is SearchIndex.from_existing(name, redis_url=...).

• connect(...) and set_client(...) are async.

• paginate is an async generator: async for batch in index.paginate(query, page_size=N):.

• No context-manager support equivalent to SearchIndex.__enter__ / __exit__ — manage lifetime manually with await index.disconnect().

Vectorizers expose .aembed(text) and .aembed_many(texts) for the embedding call itself — pair them with AsyncSearchIndex for end-to-end async pipelines (see §10).

13. LLM primitives (summary level)

Full coverage of the LLM-primitive surface is deferred to a future dedicated spec. The summaries below are intentionally narrow — minimal constructor + the upstream notebook to read for depth. None of these primitives are necessary for the core search/index/query surface this reference covers.

SemanticCache — semantic prompt → response cache

Mirrors 03_llmcache.ipynb cell 5.

from redisvl.extensions.cache.llm import SemanticCache
from redisvl.utils.vectorize import HFTextVectorizer

llmcache = SemanticCache(
    name="llmcache",                                              # underlying search index name
    redis_url="redis://localhost:6379",
    distance_threshold=0.1,                                       # cosine distance [0, 2]; lower = stricter
    vectorizer=HFTextVectorizer("redis/langcache-embed-v2"),
)

llmcache.store(prompt="What is the capital of France?", response="Paris")
hit = llmcache.check(prompt="capital city of France?")          # returns cached response on semantic match

filterable_fields=[{"name": "user_id", "type": "tag"}] partitions the cache by tenant/user. Defer to 03_llmcache.ipynb for filterable-field semantics, TTLs, and metadata.

LangCache vs SemanticCache: SemanticCache is the in-process Python class shown above. LangCache (13_langcache_semantic_cache.ipynb) is a separate Redis-hosted product — a managed semantic-cache service on Redis Cloud. They share a vector-search shape but live in different scopes. For LangCache coverage, fall through to the redis-semantic-cache skill (a separate skill in this repo) rather than treating it inline as a RedisVL primitive.

MessageHistory — durable chat-history with optional semantic recall

Mirrors 07_message_history.ipynb cells 1, 12.

from redisvl.extensions.message_history import MessageHistory, SemanticMessageHistory

# Plain FIFO chat history.
chat = MessageHistory(name="student tutor")
chat.add_message({"role": "user", "content": "Explain backprop."})
recent = chat.get_recent(top_k=8)

# Vector-recall over the same history — fetches semantically similar past turns.
semantic = SemanticMessageHistory(name="tutor")
semantic.add_messages(recent)
relevant = semantic.get_relevant("How does gradient descent relate to backprop?")

Defer to 07_message_history.ipynb for session tagging, role filtering, TTLs.

SemanticRouter — route a query to a labelled bucket via vector match

Mirrors 08_semantic_router.ipynb cells 2, 4.

from redisvl.extensions.router import SemanticRouter, Route
from redisvl.utils.vectorize import HFTextVectorizer

tech = Route(
    name="technology",
    references=["what are the latest advancements in AI?", "tell me about the newest gadgets"],
    metadata={"category": "tech"},
    distance_threshold=0.71,
)

router = SemanticRouter(
    name="topic-router",
    vectorizer=HFTextVectorizer(),
    routes=[tech, ...],
    redis_url="redis://localhost:6379",
    overwrite=True,
)

match = router("what's new in machine learning?")           # -> Route name + score

Defer to 08_semantic_router.ipynb for multi-route disambiguation, from_dict / from_yaml persistence, and threshold tuning.

Other extensions (also summary-only)

• EmbeddingsCache (redisvl.extensions.cache.embeddings.EmbeddingsCache) — cache for embed() calls to avoid recomputation. See 10_embeddings_cache.ipynb.

• Rerankers (redisvl.utils.rerank) — HFCrossEncoderReranker, CohereReranker, VoyageAIReranker for second-stage cross-encoder reranking after initial vector retrieval. See 06_rerankers.ipynb.

14. Common errors & version gotchas

redis-py equivalent: see python-redis-py.md#13-common-errors--version-gotchas.

DIALECT defaults: RedisVL passes dialect=2 on every query class by default. You do not need to set it. Override to dialect=3 for GEOSHAPE WITHIN/CONTAINS. The Redis server default (DIALECT 2 from Redis 8.0+) does not affect RedisVL because RedisVL always sends DIALECT explicitly.

rvl CLI tool: the rvl CLI (rvl index info, rvl stats, rvl index list) is a productivity tool for inspecting indexes from the shell. Deferred from v1 of this reference — agents generating Python code rarely need it. See cli.ipynb upstream.

15. Upstream examples index

redis-py equivalent: see python-redis-py.md#14-upstream-examples-index.

Curated mapping of operation → notebook + cell. Upstream files live at https://github.com/redis/redis-vl-python/blob/main/docs/user_guide/<file>. Source classes live in redis/redis-vl-python/redisvl/....