Venice errors & retries

Every Venice endpoint returns one of four error shapes. Knowing which shape you got tells you how to react.

Error body shapes

1. StandardError — simple message

The default shape for 4xx/5xx. Emitted when there's nothing structured to surface.

{ "error": "Unauthorized" }

2. DetailedError — Zod validation failure

Used for some 400 responses on malformed request bodies. When present, details is a Zod format() tree (_errors recursively keyed by field) alongside a flat issues array. Many 400s are plain StandardError without details — always handle both.

{
  "error": "Invalid request",
  "details": {
    "_errors": [],
    "messages": { "_errors": ["Field is required"] }
  },
  "issues": [
    { "code": "invalid_type", "path": ["messages"], "message": "Field is required" }
  ]
}

Render details / issues to the user so they can fix the input; don't retry — the request shape is wrong.

3. ContentViolationError — 422 content policy

Returned when a prompt trips content policy. suggested_prompt (a model-provided safe alternative) is currently emitted by the audio generation pipeline (/audio/queue, /audio/retrieve); image and video endpoints return { error: "Content policy violation" } without suggested_prompt.

{
  "error": "Content policy violation",
  "suggested_prompt": "A cinematic instrumental track inspired by stormy weather and dramatic tension."
}

Pattern — when suggested_prompt is present, retry once with prompt = suggested_prompt if the user consents.

4. X402InferencePaymentRequired — 402 on x402 inference calls

Returned only when the caller authenticated with SIWE and has insufficient credit. Discriminated by code: "PAYMENT_REQUIRED".

{
  "error": "Payment required",
  "code": "PAYMENT_REQUIRED",
  "message": "Insufficient x402 balance",
  "suggestedTopUpUsd": 10,
  "minimumTopUpUsd": 5,
  "supportedTokens": ["USDC"],
  "supportedChains": ["base"],
  "topUpInstructions": {
    "step1": "POST /api/v1/x402/top-up with no payment header to get payment requirements",
    "step2": "Sign a USDC transfer authorization using the x402 SDK (createPaymentHeader)",
    "step3": "POST /api/v1/x402/top-up with the signed X-402-Payment header",
    "receiverWallet": "<RECEIVER_WALLET_ADDRESS>",
    "tokenAddress": "<USDC_TOKEN_ADDRESS>",
    "tokenDecimals": 6,
    "network": "eip155:8453",
    "minimumAmountUsd": 5
  },
  "siwxChallenge": { ... SIWE template ... }
}

The PAYMENT-REQUIRED response header carries a base64-encoded x402 v2 paymentRequired object (x402Version, error, resource, accepts[], optional extensions) — it is not the same JSON as the body. Protocol-level clients parse the header; human-facing clients parse the richer body. See venice-x402.

Status code map

Rate-limit headers (429)

Emitted on /crypto/rpc/{network}:

Additionally, LlmInferenceError model-overloaded conditions set a Retry-After header (seconds) on the 429 — honor it when present.

Inference endpoints (chat, image, audio, video) use a per-API-key tier defined via /api_keys/rate_limits. See venice-api-keys to pre-fetch your caps, and venice-billing for DIEM/USD usage.

Response headers on 402 (x402)

Retry strategy

Never retry

• 400 — bad input. Fix the request.
• 401 — bad auth. Fix credentials.
• 403 — not entitled. Don't hammer.
• 415 — wrong Content-Type.

Retry with modification

• 402 (x402) — run top-up then retry.
• 402 (Bearer) — surface to user; top up at venice.ai.
• 422 with suggested_prompt — one retry with the safer prompt.

Retry with backoff

• 429 — back off for at least X-RateLimit-Reset - now(). Add jitter.
• 500 / 503 / 504 — exponential backoff (e.g. 0.5s, 1s, 2s, 4s, 8s), capped at ~30s. 3–5 retries max.
• Use Idempotency-Key (e.g. on /crypto/rpc/{network}) so retries can't double-bill state-mutating calls.

Reference retry loop

async function callVenice<T>(fn: () => Promise<Response>): Promise<T> {
  const maxRetries = 5
  let delay = 500
  for (let attempt = 0; attempt <= maxRetries; attempt++) {
    const res = await fn()
    if (res.ok) return res.json() as Promise<T>

    const body = await res.clone().json().catch(() => ({}))
    const { status } = res

    if ([400, 401, 403, 415].includes(status)) {
      throw Object.assign(new Error(body.error ?? 'Venice error'), { status, body })
    }

    if (status === 402 && body.code === 'PAYMENT_REQUIRED') {
      await topUpX402(body.suggestedTopUpUsd)
      continue
    }

    if (status === 422) {
      throw Object.assign(new Error('Content policy'), { status, body })
    }

    if (status === 429) {
      const retryAfterSec = Number(res.headers.get('retry-after'))
      const resetSec = Number(res.headers.get('x-ratelimit-reset'))
      const waitMs = !Number.isNaN(retryAfterSec) && retryAfterSec > 0
        ? retryAfterSec * 1000
        : !Number.isNaN(resetSec) && resetSec > 0
          ? Math.max(resetSec * 1000 - Date.now(), delay)
          : delay
      await sleep(waitMs + Math.random() * 250)
      delay *= 2
      continue
    }

    if (status >= 500 && attempt < maxRetries) {
      await sleep(delay + Math.random() * 250)
      delay *= 2
      continue
    }

    throw Object.assign(new Error(body.error ?? 'Venice error'), { status, body })
  }
  throw new Error('Exceeded max retries')
}

Streaming errors

Streaming responses (stream: true on chat, TTS, video-queue progress) deliver mid-stream errors as SSE events:

data: {"error": {"type": "…", "message": "…"}}

Treat them as terminal — the underlying connection is closed. The HTTP status is 200 because a successful stream can't be changed mid-flight.

Request-ID correlation

When present on a response, keep the X-Request-ID header. Include it in support tickets — Venice keys diagnostic logs by this ID. /crypto/rpc/* routes set it explicitly; many inference routes also include it, but don't assume it's universal — fall back to your own client-side correlation ID.

Common gotchas

• A 402 from /x402/top-up with no X-402-Payment header is the expected discovery response, not an error. See venice-x402.
• A 500 on /chat/completions with a huge file upload often means the upstream model chose to abort — reduce max_tokens / image size rather than blindly retrying.
• 429 on /crypto/rpc/{network} may mean the 24-hour credit cap tripped, not the per-minute one. Check customMessage.
• DetailedError.details is a Zod _errors tree, not a flat map. Walk it recursively.
• Some endpoints (image generation) echo X-Rate-Limit variants — treat any header whose name starts with X-RateLimit as advisory.
• Don't treat an empty stream chunk as an error — send-keepalives look like data: [DONE] or empty lines.