venice-embeddings

Venice Embeddings

POST /api/v1/embeddings returns vector embeddings for strings. It's OpenAI-compatible: request and response match https://api.openai.com/v1/embeddings closely enough that the OpenAI SDK works with baseURL: "https://api.venice.ai/api/v1", with one exception: token-ID arrays are not accepted (see below).

Auth: Bearer API key or x402 wallet (SIGN-IN-WITH-X) — see venice-auth.

Use when

• You're building retrieval / RAG / similarity search.
• You need text clustering, classification, deduplication, or reranking features.
• You want embeddings from a model Venice runs as Private (most of the catalog) rather than an anonymized third-party one — check model_spec.privacy on each model.

Text-only: input must be a string or an array of strings. For images, run them through a vision chat model and embed the description.

Minimal request

curl https://api.venice.ai/api/v1/embeddings \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  --compressed \
  -d '{
    "model": "text-embedding-bge-m3",
    "input": "Why is the sky blue?"
  }'

{
  "object": "list",
  "model": "text-embedding-bge-m3",
  "data": [
    { "object": "embedding", "index": 0, "embedding": [0.0023, -0.0093, 0.0158, ...] }
  ],
  "usage": { "prompt_tokens": 8, "total_tokens": 8 }
}

Request schema

The body is strict — unknown top-level fields (e.g. input_type, task, truncate) are rejected with 400.

Field	Type	Notes
model	string	Required. Model ID from GET /models?type=embedding. text-embedding-ada-002 is accepted as an alias for text-embedding-bge-m3 (the response model then reads text-embedding-bge-m3).
input	string | string[]	Required. A non-empty string, or an array of 1–2048 strings. Token-ID arrays (number[] / number[][]) are rejected with 400 — see Gotchas.
encoding_format	"float" | "base64"	Default "float". "base64" returns each vector as a base64-encoded string — a much smaller payload; decode client-side.
dimensions	integer ≥ 1	Optional. Requested output size. Only honoured by models whose model_spec.supportsCustomDimensions is true; it is forwarded as-is otherwise, and the provider may ignore or reject it.
user	string	Accepted for OpenAI compatibility; not used for inference, but it does split the error budget per value (see venice-errors).

Input limits

• Per item: Venice estimates tokens as ceil(chars / 3.2 × 0.95) and rejects any string over 8192 estimated tokens (≈ 27,500 characters) with 400 "Input text exceeds the maximum token limit of 8192 tokens". This cap is the same for every model, including those whose maxInputTokens is 32768.
• Model limit: models with a smaller model_spec.maxInputTokens (e.g. 512 for text-embedding-multilingual-e5-large-instruct, 2048 for gemini-embedding-2-preview) enforce that limit themselves; Venice's pre-check does not. Chunk to the model's maxInputTokens.
• Batch: at most 2048 strings per request. Venice returns one embedding per element, in order, with matching index.

Response headers & compression

Send Accept-Encoding: gzip, br (curl: --compressed; most HTTP clients decode automatically); the response comes back with Content-Encoding set. For large batches this matters — float vectors in JSON are big.

Also returned:

• x-ratelimit-limit-* / x-ratelimit-remaining-* / x-ratelimit-reset-* (requests, tokens) — see venice-errors.
• x-venice-balance-usd / x-venice-balance-diem — current balance, when non-zero.
• X-Balance-Remaining — listed in the spec for x402 callers but not currently set by the server; poll GET /x402/balance/{walletAddress} instead.

Using the OpenAI SDK

import OpenAI from 'openai'

const client = new OpenAI({
  apiKey: process.env.VENICE_API_KEY,
  baseURL: 'https://api.venice.ai/api/v1',
})

const res = await client.embeddings.create({
  model: 'text-embedding-bge-m3',
  input: ['first doc', 'second doc'],
})

const vec0 = res.data[0].embedding

The OpenAI Node SDK asks for base64 when you omit encoding_format and decodes it client-side. Venice forwards encoding_format to the model; if the decoded vectors look wrong, pass encoding_format: 'float' explicitly.

LangChain

LangChain's OpenAIEmbeddings tokenizes input and sends token arrays by default, which Venice rejects. Turn that off:

import os
from langchain_openai import OpenAIEmbeddings

emb = OpenAIEmbeddings(
    model="text-embedding-bge-m3",
    base_url="https://api.venice.ai/api/v1",
    api_key=os.environ["VENICE_API_KEY"],
    check_embedding_ctx_length=False,
)

Wrappers that build OpenAIEmbeddings internally without that flag (e.g. gpt-researcher's openai provider) hit the same 400.

Batch-embedding pattern

async function embedBatch(texts: string[], batchSize = 64) {
  const out: number[][] = []
  for (let i = 0; i < texts.length; i += batchSize) {
    const slice = texts.slice(i, i + batchSize)
    const res = await client.embeddings.create({
      model: 'text-embedding-bge-m3',
      input: slice,
      encoding_format: 'float',
    })
    for (const row of res.data) out[i + row.index] = row.embedding
  }
  return out
}

• Keep each string under the per-item limits above; batch size is capped at 2048.
• On 429, back off exponentially and halve the batch — see venice-errors.

Choosing a model

Query GET /models?type=embedding for the current catalog. Each entry's model_spec exposes:

• embeddingDimensions — native output dimension (e.g. 1024 for text-embedding-bge-m3, 4096 for text-embedding-qwen3-8b).
• maxInputTokens — the model's per-input token limit.
• supportsCustomDimensions — present and true only on models that honour dimensions (absent otherwise).
• privacy — "private" (no retention) or "anonymized" (a third-party model; the request is sent without your identity).
• pricing.input / pricing.output — { usd, diem } per million tokens.

Representative IDs: text-embedding-bge-m3 (private, 1024-d), text-embedding-qwen3-8b (private, 4096-d, custom dimensions), text-embedding-multilingual-e5-large-instruct (private, 512-token inputs), text-embedding-3-small / text-embedding-3-large (anonymized, OpenAI, custom dimensions), gemini-embedding-2-preview (anonymized, custom dimensions). The list changes — always read it from /models.

Always pin the model ID — cosine distances are not comparable across different embedding models.

Error handling

Code	Meaning
400	Missing model, or a validation error (details names the field): token-array input, empty string/array, > 2048 items, item over the 8192-token estimate, unknown field. Non-JSON Content-Type → "'Content-Type' must be 'application/json'". Also model-side rejections (e.g. input over the model's own limit), returned with the model's message when one can be extracted, otherwise a generic "Invalid request parameters…".
401	Invalid API key or SIWX signature.
402	Insufficient balance or the key's USD/DIEM spend limit reached. Bearer → "Insufficient USD or Diem balance…"; x402 → payment-required body + PAYMENT-REQUIRED header. A request with no credentials at all also gets 402 (x402 discovery challenge), not 401.
403	The API key's modelPrivacy is PRIVATE_TEXT or PRIVATE_ONLY and the model is anonymized. Also region / provider restrictions, or API access disabled for the account.
404	Unknown model (the message may suggest a close match).
429	Rate limited.
500	Inference failed; retry with jitter.
503	Model temporarily offline; retry later.

Gotchas

• Token arrays are rejected. input: [101, 2023, ...] or [[101, ...]] returns 400 "Token array inputs are not supported. Pass a string or an array of strings." Send text.
• API-key privacy applies to embeddings. A key with modelPrivacy: PRIVATE_TEXT (or PRIVATE_ONLY) can only call private embedding models; text-embedding-3-* and gemini-embedding-2-preview return 403. See venice-api-keys.
• A top-level empty string is rejected with 400, but empty strings inside an array are not pre-checked — filter them out yourself.
• The request must be Content-Type: application/json; anything else is rejected with 400 "'Content-Type' must be 'application/json'" before auth or validation runs.
• Whether returned vectors are L2-normalized depends on the model — verify with Math.hypot(...v) ≈ 1 before assuming.
• For RAG, store model (and dimensions, if set) alongside each vector so you can re-embed on upgrade.