API reference

The presented API key's identity, scopes, and quota

curl -H "Authorization: Bearer $KI_API_KEY" \
  "https://kernelindex.com/api/v1/me"

Resolver decision for a text query

q (query) · cohort (query)

curl "https://kernelindex.com/api/v1/search?q=rmsnorm%20b200%20bf16"

Response shape; values elided.

{
  "query": "rmsnorm b200 bf16",
  "mode": "exact",
  "operation": { "name": "…", "slug": "…" },
  "policyVersion": "ranking-v1",
  "cohort": { "comparisonKey": "sha256:…", "profile": "…", "facts": [ … ] },
  "cohortOptions": [ { "key": "sha256:…", "label": "NVIDIA B200", "runs": …, "head": { … } } ],
  "bestVerified": null,
  "bestDeployable": { "runId": "…", "implementation": { … }, "primary": { "metric": "latency", "unit": "ns", "value": … },
                      "rank": …, "cohortSize": …, "install": { "kind": "pip", "command": "…" }, "evidence": "reported", … },
  "groups": { "exact": [ … ], "compatible": [ … ], "supportedUnmeasured": [ … ], "reported": [ … ] },
  "overflow": { "exact": 0, "compatible": 0, "supportedUnmeasured": 0, "reported": 0 },
  "nearest": null,
  "sources": [ { "name": "…", "observedAt": "…" } ],
  "generatedAt": "…"
}

Resolver decision for a structured request

curl -X POST "https://kernelindex.com/api/v1/resolve/kernel" \
  -H "Content-Type: application/json" \
  -d '{"operation": {"name": "rmsnorm"}, "environment": {"hardwareProduct": "B200", "dtype": "bf16"}}'

One resolver decision per request, in request order

curl -X POST "https://kernelindex.com/api/v1/resolve/kernel/batch" \
  -H "Content-Type: application/json" \
  -d '{"requests": [
    {"operation": {"name": "rmsnorm", "axes": {"tokens": 2048}},
     "environment": {"hardwareProduct": "B200", "dtype": "bf16"}},
    {"operation": {"name": "gemm"}, "environment": {"hardwareProduct": "B200"}}
  ]}'

Implementations most likely to contain transferable optimization ideas for the requested kernel problem, ranked by transferability (same computation, same or adjacent hardware, adjacent shape, proven standing, shared techniques), each with its reasons. Not a compatibility answer: the problem need not be indexed.

curl -X POST "https://kernelindex.com/api/v1/precedents" \
  -H "Content-Type: application/json" \
  -d '{}'

Operation dossier (the web page's model)

idOrSlug (path, required) · workload (query) · cohort (query)

curl "https://kernelindex.com/api/v1/operations/<idOrSlug>"

Implementation dossier

idOrSlug (path, required) · include (query)

curl "https://kernelindex.com/api/v1/implementations/<idOrSlug>"

Project dossier: standing, measured implementations, claim state

slug (path, required)

curl "https://kernelindex.com/api/v1/projects/<slug>"

Immutable run evidence dossier

idOrDigest (path, required)

curl "https://kernelindex.com/api/v1/runs/<idOrDigest>"

idOrDigest (path, required)

curl -X POST "https://kernelindex.com/api/v1/runs/<idOrDigest>/attestations" \
  -H "Content-Type: application/json" \
  -d '{}'

Records ledger page (cursor-paginated)

cursor (query) · limit (query)

curl "https://kernelindex.com/api/v1/records"

Aligned comparison of 2–8 runs

curl -X POST "https://kernelindex.com/api/v1/compare" \
  -H "Content-Type: application/json" \
  -d '{"runs": ["<run-id>", "<run-id>"]}'

Redirect to the latest immutable catalog export

curl "https://kernelindex.com/api/v1/exports/catalog.jsonl.zst"

Validation report plus, per run, the cohort it would join and the rank it would take under ranking-v1. Never a promise; review decides comparability

curl -X POST "https://kernelindex.com/api/v1/submissions/preview" \
  -H "Content-Type: application/json" \
  -d '{}'

curl -X POST "https://kernelindex.com/api/v1/submissions" \
  -H "Content-Type: application/json" \
  -d '{}'

Retract a run or mark it superseded (site_admin session required).

curl -X POST "https://kernelindex.com/api/v1/corrections" \
  -H "Content-Type: application/json" \
  -d '{"action": "retract", "runId": "<run-id>", "reason": "…"}'

Drop catalog caches (importer bearer token required).

curl -X POST "https://kernelindex.com/api/v1/revalidate" \
  -H "Content-Type: application/json" \
  -d '{}'

Feasible serving configurations grouped by cohort, ranked only under an explicit objective; the Pareto frontier otherwise

curl -X POST "https://kernelindex.com/api/v1/resolve/serving" \
  -H "Content-Type: application/json" \
  -d '{"model": "llama2-70b-99", "objective": {"direction": "maximize", "metric": "output_token_throughput_tps"}}'

Serving runs, cursor-paginated

cursor (query) · limit (query)

curl "https://kernelindex.com/api/v1/serving-runs"

Serving configurations with eligible run counts

curl "https://kernelindex.com/api/v1/serving-configurations"

Serving run evidence dossier

id (path, required)

curl "https://kernelindex.com/api/v1/serving-runs/<id>"

Published runs, newest observation first, keyset-paginated

operation (query) · hardware (query) · source (query) · status (query) · since (query) · cursor (query) · limit (query)

curl "https://kernelindex.com/api/v1/runs"

Requested workloads, priority and model gaps, unbeaten baselines, unchallenged and stale records; every row points at the cohort or search where the answer would go

curl "https://kernelindex.com/api/v1/challenges"

Record breaks, publication batches, corrections, and accepted claims over the trailing 30 days, newest first, grouped by UTC day

since (query)

curl "https://kernelindex.com/api/v1/feed"

Every operation with taxonomy tags, workload count, and eligible run count

family (query) · tag (query)

curl "https://kernelindex.com/api/v1/operations"

Per-GPU coverage: kernel and serving run counts and operation-family breadth (counts, never a shared ranking)

curl "https://kernelindex.com/api/v1/hardware"

Model coverage: serving model revisions and kernel-side model: tags, as separate arrays

curl "https://kernelindex.com/api/v1/models"

Model dossier: best known per operation on one GPU, gaps, evidence, source links (the web page's model)

slug (path, required) · gpu (query)

curl "https://kernelindex.com/api/v1/models/<slug>"

Live per-source corpus counts and the hero family/GPU coverage grid

curl "https://kernelindex.com/api/v1/coverage"

Source provenance: slug, kind, run counts, and last snapshot fetch time

curl "https://kernelindex.com/api/v1/sources"