Chapter 10 · Vector Sets, Vector Search, Hybrid Retrieval, and AI Workloads
Vector Compression/Quantization Awareness, Memory Planning, and Accuracy Benchmarks
Benchmark quantization, dimension reduction, graph memory, recall, and tail latency as one coupled capacity problem.
Learning outcomes
AtlasMart's prototype retrieves good neighbors, but production capacity depends on millions of dimensions, encoded vectors, graph links, attributes, persistence, and concurrency. Memory optimization is only useful if retrieval quality remains acceptable.
Compare Q8, BIN, and NOQUANT as measurable Vector Set representation choices.
Estimate raw coordinate memory and separate it from graph/label/attribute overhead.
Explain how M, EF, REDUCE, and dimension affect memory, ingestion, latency, and recall.
Build an exact ground-truth benchmark with VSIM TRUTH and compute recall@k.
Produce p50/p95/p99 latency plus memory-per-vector evidence without fabricated production numbers.
All Chapter 10 mandatory labs reuse the disposable Chapter 01
environment: Redis Open Source 8.10.1 from Docker
Official Image redis:8.10.1, container
atlasmart-redis-ch01, standalone topology, host
publication 127.0.0.1:6379, TLS disabled only
because traffic stays on loopback, default ACL user disabled,
named users atlasmart-app and
academy-admin, logical database 0, AOF with
appendfsync everysec plus RDB snapshots,
persistent /data, and no explicit
maxmemory limit or eviction policy. Redis 8
integrates Vector Sets and the Redis Query Engine into Redis
Open Source. Mandatory examples use synthetic numeric vectors
created locally—Redis stores/searches vectors but does not
generate embeddings. Fixtures stay under
atlasmart:ch10:*. Vector Set commands use the
restricted application user where allowed; Search index
administration uses the disposable
academy-admin user. No paid embedding API,
managed service, production endpoint, or real credential is
required.
Redis 8.0 introduced Vector Sets as a beta data type. The
current Redis Open Source 8.10 command reference documents
VADD, VSIM, VINFO,
filtering, quantization, and related commands as available
since 8.0, with standard Redis Software/Redis Cloud
compatibility. The official sources checked for this lesson do
not provide a separate explicit “Vector Sets became GA on
version X” declaration. Treat Vector Set API/product status,
client coverage, managed-service support, and Active-Active
compatibility as version-sensitive and verify the exact target
rather than inventing a GA date.
1. Start with raw coordinate arithmetic, then measure Redis
A 300-dimensional FP32 vector contains 300 × 4 = 1200 raw coordinate bytes before labels, object headers, allocator effects, HNSW links, attributes, and persistence buffers. Q8 uses approximately one byte/component (about 4× smaller than FP32 coordinates); BIN is approximately one bit/component (about 32× smaller). These are coordinate-level ratios, not complete key memory.
| Mode | Coordinate intuition | Quality expectation |
|---|---|---|
| NOQUANT | full FP32-style coordinate footprint | highest representation fidelity, highest memory |
| Q8 | ~1 byte/component; default | high recall/efficiency balance |
| BIN | ~1 bit/component | lowest coordinate memory, lower recall |
2. Measure whole-key memory with MEMORY USAGE
docker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app DEL atlasmart:ch10:bench:q8 atlasmart:ch10:bench:noq atlasmart:ch10:bench:bindocker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app VADD atlasmart:ch10:bench:q8 VALUES 4 1.262185 1.958231 0.4 0.9 item Q8docker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app VADD atlasmart:ch10:bench:noq VALUES 4 1.262185 1.958231 0.4 0.9 item NOQUANTdocker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app VADD atlasmart:ch10:bench:bin VALUES 4 1.262185 1.958231 0.4 0.9 item BINdocker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app MEMORY USAGE atlasmart:ch10:bench:q8docker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app MEMORY USAGE atlasmart:ch10:bench:noqdocker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app MEMORY USAGE atlasmart:ch10:bench:binA one-vector fixture is too small for production capacity inference because fixed overhead dominates. Repeat at representative N/dimension and divide measured whole-key memory by cardinality while also reporting fixed/key-level overhead.
3. Quantization changes vectors and may change neighbors
docker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app VEMB atlasmart:ch10:bench:q8 itemdocker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app VEMB atlasmart:ch10:bench:noq itemdocker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app VEMB atlasmart:ch10:bench:bin itemBIN can visibly collapse coordinate detail. The correct question is not “which representation looks closest?” but “does the representation preserve the required top-k neighbors and task quality under the target workload?”
4. Exact truth makes recall measurable
For each benchmark query, run an exact
VSIM ... TRUTH result and an approximate result
with identical COUNT k. Compute recall@k. Repeat
across queries, not one cherry-picked vector.
def recall_at_k(exact, approx, k): return len(set(exact[:k]) & set(approx[:k])) / kqueries=[ (["a","b","c","d"],["a","b","c","x"]), (["m","n","o","p"],["m","n","q","p"]),]vals=[recall_at_k(e,a,4) for e,a in queries]print(vals, sum(vals)/len(vals))5. M buys graph connectivity with memory
M controls maximum neighbor links in the HNSW
graph. Current Redis Vector Set docs note that layer 0 uses
roughly 2*M links and higher layers roughly
M, with pointer memory contributing significantly.
Raising M without a recall problem wastes memory and ingestion
work.
Budget coordinates + graph links + labels + attributes + key/allocator overhead + persistence/replication/fork headroom. maxmemory is not the same as process RSS.
6. REDUCE changes geometry, not just bytes
Random projection through REDUCE lowers stored
dimension and saves coordinate/graph-adjacent work, but it
changes distances. Benchmark reduced and unreduced systems on
the same query/relevance set.
docker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app DEL atlasmart:ch10:bench:full atlasmart:ch10:bench:reduceddocker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app VADD atlasmart:ch10:bench:full VALUES 4 1 0.2 0.1 0.7 item-adocker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app VADD atlasmart:ch10:bench:reduced REDUCE 2 VALUES 4 1 0.2 0.1 0.7 item-adocker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app VDIM atlasmart:ch10:bench:fulldocker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app VDIM atlasmart:ch10:bench:reduced7. Search vector indexes have a separate compression design space
Redis Search vector fields support FLAT, HNSW, and SVS-VAMANA. Search also supports multiple numeric vector types and SVS-oriented compression options depending on release. Do not map Vector Set Q8/BIN knobs mechanically onto Search index settings; they are different APIs and representations.
| System | Exact option | Approximate option | Compression/representation |
|---|---|---|---|
| Vector Set | VSIM TRUTH baseline | built-in HNSW-style VSIM | Q8 default, BIN, NOQUANT, REDUCE |
| Search vector field | FLAT | HNSW / SVS-VAMANA | vector TYPE plus algorithm-specific compression options |
8. Latency benchmark contract
For every configuration record server/patch, CPU/RAM, container/native, dimension, N, quantization, M/build-EF/query-EF, filter selectivity, query concurrency, warmup, persistence/fsync, pipeline/client connection model, and p50/p95/p99. Run enough samples for stable tails. Do not paste the demonstration numbers below into capacity documents.
# Replace with timings captured from your benchmark harness.samples_ms=[1.02,1.07,1.08,1.11,1.15,1.18,1.24,1.31,1.52,2.10]def nearest_rank(xs,p): xs=sorted(xs); return xs[max(0,min(len(xs)-1, math.ceil(p*len(xs))-1))]import mathfor p in (0.50,0.95,0.99): print(p, nearest_rank(samples_ms,p))9. Memory-per-vector benchmark contract
Measure MEMORY USAGE after a warm, fully built
fixture and VCARD. Report both total bytes and
bytes/vector. For attributes, test representative lengths and
selectivities. Account for temporary ingestion/build peaks and
fork/AOF/replication headroom separately.
10. Wrong approach: publish “Q8 is 4× cheaper” as total Redis memory
The 4× figure refers primarily to coordinate representation versus FP32; graph links, labels, attributes, allocator metadata, and fixed key overhead do not shrink by the same factor. Repair by measuring whole-key memory at realistic cardinality and disclosing what is included.
11. Wrong approach: optimize latency with BIN and never re-check relevance
Binary quantization may be fast and memory-efficient while changing nearest-neighbor order. Any representation/tuning change that affects geometry requires regression on exact recall@k and downstream task metrics.
12. Reproducible cleanup
docker exec -e REDISCLI_AUTH=AtlasMart-App-Lab-Only-2026 atlasmart-redis-ch01 redis-cli --user atlasmart-app DEL atlasmart:ch10:bench:q8 atlasmart:ch10:bench:noq atlasmart:ch10:bench:bin atlasmart:ch10:bench:full atlasmart:ch10:bench:reduced13. Production judgment
Capacity planning couples memory, recall, and latency. Keep enough headroom for replication/AOF buffers and fork-based persistence; large Vector Sets can become hot keys on a standalone/Cluster slot; retries and background search threads affect tails; patch releases can contain Vector Set correctness/security fixes. Rebuild/migration plans must include source embeddings or a reproducible embedding pipeline because a highly quantized in-memory representation is not necessarily the right long-term source artifact.
14. Summary and next step
You can now benchmark representation choices instead of arguing from folklore. Lesson 5 assembles the pieces into production-shaped semantic search, recommendation, and retrieval-augmented generation (RAG) designs with evaluation and security gates.
Check your understanding
- Why is Q8 “4× smaller” not a whole-key memory guarantee?
- What is VSIM TRUTH for?
- What does M change?
- Why does REDUCE require a recall regression?
- Which latency percentiles should the prompt explicitly measure?
Review the answers
Graph, labels, attributes, allocator/key overhead do not shrink by the same coordinate ratio.
Exact linear-scan ground truth and recall benchmarking.
HNSW graph connectivity, affecting memory/build/search behavior.
It changes vector geometry through projection.
At least p50, p95, and p99 for this chapter.
Authoritative references
- Redis Vector Sets — native Vector Set data type, commands, filtering, and examples
- VADD — Vector Set insertion, quantization, REDUCE, EF, M, and attributes
- VSIM — similarity queries, scores, filters, EF, TRUTH, and NOTHREAD
- VINFO — Vector Set implementation and configuration evidence
- VEMB — stored/reconstructed vector evidence
- VSETATTR — JSON attributes attached to Vector Set elements
- Vector Set memory optimization — Q8/BIN/NOQUANT, dimensions, graph links, and memory tradeoffs
- Vector Set performance — quantization and vector-set performance considerations
- Redis vector search concepts — Search FLAT/HNSW/SVS-VAMANA vector indexes and runtime parameters
- Vector field options — Search vector field types, metrics, and index algorithms
- FT.CREATE — Search schema and vector field creation
- FT.SEARCH — KNN/hybrid query syntax and parameters
- Redis 8.10 commands — target-version command surface
- Redis 8.10 release notes — 8.10.1 security baseline including Vector Set fixes
- Redis 8 GA announcement — historical 8.0 Vector Set beta status and integrated Redis 8 capabilities