llm·yard

Sections 4 and 5 · Pin configuration and electrical characteristics.

Pin configuration

Bearer is the per-project virtual key on all of these.
PinFunctionType
1/v1/chat/completionsfull loop
2/v1/messagesanthropic wire
3/v1/embeddingspassthrough
4/v1/responsespassthrough
5/v1/rerankpassthrough
6/v1/moderationspassthrough
PinFunctionType
7/v1/images/generationsadapter
8/v1/audio/speechadapter
9/video/generationssubmit, poll, fetch
10/v1/modelsread, validates key
11/status/peekread, $0 pre-call gate
12/usage/meread, own spend

Speech to text has no pin. That caller still goes to its vendor directly. Two headers a caller may send on any surface: x-axon-source names the calling application, for routing, limits and attribution; x-axon-timeout-ms bounds the whole failover chain rather than one attempt, so a chain of N entries cannot keep a caller waiting N times the per-attempt timeout.

Absolute maximum ratings

Operating a part beyond these is refused, not degraded.
ParameterSymbolValue
Request bodyAXON_MAX_BODY_MB32 MB
Upstream bodyAXON_MAX_UPSTREAM_BODY_MB64 MB
Caller deadline ceilingAXON_MAX_REQUEST_TIMEOUT_MS120 000 ms
Stream read gapAXON_STREAM_READ_TIMEOUT_SECS300 s
ParameterSymbolValue
Captured body, per sideaxon_request_bodies32 KB
Injection screen, per fieldguardrails1 MiB
Default calls per key…KEY_CALLS_PER_HOUR1,000 /h
Default cost per key…KEY_COST_PER_DAY_CENTS$5.00 /day

Electrical characteristics

Shipped defaults, at nominal conditions.
ParameterSymbolMinTypMaxUnit
Semantic cache matchCACHE_SIMILARITY0.951.00cosine
Cache entry lifeCACHE_TTL_SECS3600s
Embedding widthCACHE_EMBED_DIM102410241024dim
Config convergenceCONFIG_REFRESH_SECS30s
Model catalog refreshMODEL_REFRESH_SECS21600s
Failover price guard…FAILOVER_MAX_PRICE_MULT3.0×
Soft cap downrouteAXON_SOFTCAP_PCT90100%

Every variable the code reads appears in .env.example, grouped by purpose, with its real default stated. llmyard does not parse a .env file itself; Docker Compose loads it and passes it in.

Vendor presets

9 vendors, one credential each.

Each preset carries a documented base URL, the auth style the vendor wants, the wire dialect it speaks and a starting model list, so registering one is a credential and a click: openai, anthropic, groq, openrouter, mistral, deepseek, together, gemini, openai-embed. Anything else with an OpenAI-shaped or Anthropic-shaped endpoint works as a custom endpoint.