continuedev/continue · error · Error

Unsupported embeddings type received: number[][]

Error message

Unsupported embeddings type received: number[][]

What it means

getEmbedTexts validates the OpenAI-style embedding `input` parameter: Bedrock embedding models only accept strings (or an array of strings), not pre-tokenized token-ID arrays. If the first element of input is itself an array (i.e. input is number[][]), the adapter throws this error before any AWS call is made.

Source

Thrown at packages/openai-adapters/src/apis/Bedrock.ts:617

    const command = new InvokeModelCommand(payload);
    const client = await this.getClient();
    const response = await client.send(command);
    if (!response.body) {
      throw new Error("No response body");
    }
    const decoder = new TextDecoder();
    const decoded = decoder.decode(response.body);
    return JSON.parse(decoded);
  }

  private getEmbedTexts(body: EmbeddingCreateParams): string[] {
    const texts: string[] = [];
    if (typeof body.input === "string") {
      texts.push(body.input);
    } else if (body.input.length > 0) {
      const firstVal = body.input[0];
      if (Array.isArray(firstVal)) {
        throw new Error("Unsupported embeddings type received: number[][]");
      }
      if (typeof firstVal === "string") {
        texts.push(...(body.input as string[]));
      } else {
        throw new Error("Unsupported embeddings type received: number[]");
      }
    }
    return texts;
  }

  async embed(body: EmbeddingCreateParams): Promise<CreateEmbeddingResponse> {
    const texts = this.getEmbedTexts(body);

    let embeddings: number[][];
    if (body.model.startsWith("cohere")) {
      const payload = {
        texts,
        input_type: "search_document",

View on GitHub (pinned to 5522c6f44c)

Solutions

1. Pass plain strings: embed({model, input: ['hello world', ...]}).
2. If your pipeline holds token IDs, decode them back to text with the same tokenizer before calling embed.
3. Add a runtime type check on input before dispatching to Bedrock.

Example fix

// before
const res = await api.embed({ model: 'bedrock/titan-embed', input: [[101, 2054, 2003]] });

// after
const res = await api.embed({ model: 'bedrock/titan-embed', input: ['The answer is'] });

const isStringInput = (i: unknown): i is string | string[] => typeof i === 'string' || (Array.isArray(i) && i.every(x => typeof x === 'string'));
if (!isStringInput(body.input)) throw new TypeError('Bedrock embeddings accept only string inputs');

function isEmbeddingStringInput(input: string | string[] | number[] | number[][]): input is string | string[] {
  if (typeof input === 'string') return true;
  return Array.isArray(input) && input.every(x => typeof x === 'string');
}

When it happens

Trigger: Calling embed with input: [[1,2,3],[4,5,6]] or any array-of-arrays of token IDs; code written against OpenAI's token-embedding mode reused with the Bedrock adapter; sending token arrays produced by a tokenizer (e.g. tiktoken) directly.

Common situations: Porting OpenAI embeddings code (which accepts token arrays) to Bedrock; caching layers that pre-tokenize and store number[][]; batch pipelines that forget to decode tokens back to text.

Related errors

• Unsupported embeddings type received: number[]
• Unsupported model: ${body.model}
• Query and chunks must not be empty
• Failed to fetch messages: ${response.statusText}
• No workspace directories found

AI-assisted analysis of continuedev/continue@5522c6f44c (2026-08-27). Data as JSON: /api/errors/9cf64e6892a1cf42. Report an issue: GitHub.