The product, in detail
This page is the long answer to "what exactly do I get?", each system explained twice: once in plain words for whoever pays the bill, once in specifics for whoever deploys the binary.
All seven ship in every edition, including free Community. No other cache ships them together.
01
In plain words: It knows 'how do refunds work?' and 'what's your refund window?' are the same question, but 'cancel' and 'pause' are not.
under the hood · Every query gets an embedding in the HNSW vector index and a structural template, numbers, dates, and entities abstracted into slots. A hit requires both signals to agree, which kills the false-positive class that similarity-only caches are infamous for.
02
In plain words: One bad answer in a smart cache spreads to everyone who asks anything similar. Crowkis checks every answer's trustworthiness before storing it.
under the hood · Five weighted stages, coherence 0.30, content 0.10, source trust 0.30, tenant isolation 0.15, neighbourhood agreement 0.15, with a 0.75 composite floor. Every accept and refuse lands in an append-only trust ledger, so writers with bad history face a higher bar.
03
In plain words: The cache learns where it can afford to be generous and where it must be strict, by watching its own results.
under the hood · Per-intent reuse thresholds adjust from live hit/miss feedback within bounded ranges, with complexity adjustment and EMA decay. Twelve intent classes, each with its own bar; any threshold can be pinned via the management API.
04
In plain words: Beyond reusing answers, Crowkis reuses the way an answer was worked out, the expensive part of an LLM call.
under the hood · Chain-of-thought output is parsed into typed steps, specifics are abstracted into slots, and the step-sequence signature is stored. New inputs that match the signature get the recomposed skeleton, savings response-level caching can't reach.
05
In plain words: When space runs low, it doesn't throw out the most valuable answers, it knows what each one cost to make.
under the hood · Eviction scores recency, frequency, isolation, and compute cost at 0.25 each. A $0.40 chain-of-thought answer and a $0.0004 one-liner are not equally disposable, and the evictor knows it.
06
In plain words: Every served answer clears a quality bar first. Uncertain matches go to the model instead of guessing.
under the hood · A geometric mean of five signals, similarity, freshness, trust, hit history, intent threshold, gates every response. Factual content needs 0.88; creative gets 0.70. The geometric mean means one weak signal tanks the score, by design.
07
In plain words: Yesterday's truth doesn't outlive its shelf life. Prices change; the cache keeps up.
under the hood · Five TTL policies plus version pinning and invalidation webhooks. Freshness also feeds confidence, an aging entry decays toward recompute before it ever serves something stale.
The foundation
CrowkisDB is a purpose-built Rust LSM tree, write-ahead log with CRC-checked records, 64 MB memtable, LZ4-compressed SSTables with bloom filters, three-level compaction, with the HNSW vector index persisting beside it. No RocksDB, no garbage collector in the read path, no external dependencies.
In plain words: the part that holds your data was built for caching LLM answers specifically, which is why hits come back in under a millisecond and a power cut doesn't cost you your cache.
< 1ms
cache hits, in-process
347
integration tests on the suite
WAL
crash-safe, CRC-checked records
0 GC
Rust, no collector pauses
RESP3 Redis wire protocol
redis-py, ioredis, Lettuce connect unmodified. crowkis cli ships in the binary.
gRPC h2c, protobuf
Get / Set / GetStream / Stats / Invalidate for service meshes that prefer contracts.
REST management API
Thresholds, tenants, budgets, PII reports, compliance exports, canary control.
MCP for AI apps
Claude Code and agent frameworks check the cache before spending tokens.
Every decision leaves evidence. The dashboard and management API expose all of it.
Live verdict feed
Every hit, miss, and block streamed with its confidence score and the stage that decided it.
Cost accounting
Dollars and tokens saved, per tenant and per model, the number your CFO actually asks for.
Canary & migration
Upgrade models without torching the warm cache: canary a slice, compare, migrate with leasing.
Budgets & circuit breakers
Per-tenant spending walls enforced locally, before the invoice surprises you.
PII & compliance
Scrubbing, erasure workflows, and compliance report exports through the management API.
Prometheus & OTel
Lights up in Grafana or Datadog with zero adapters.
39 shipped & in the binary today · hover a tag for what it does and why
Identity & access·5
Multi-tenancy & isolation·3
Compliance & governance·6
Security & data protection·6
Cost, FinOps & billing·5
Reliability & operations·5
Observability & integration·4
Cache intelligence, the moat·7
Distribution & adoption·4
Identity & access
what it does · Admin, Reader, and Auditor roles, enforced on every management route.
Why it matters: The person who can flush the cache and the person who can only read reports are different people, and the server knows it.
crowkis field guide · specimen 01 of 45