FlowiseAI/Flowise · error · Error

Must specify one of "k" or "similarity_threshold".

Error message

Must specify one of "k" or "similarity_threshold".

What it means

Thrown by EmbeddingsFilterRetriever.init when both nodeData.inputs.k and nodeData.inputs.similarityThreshold are undefined. The underlying LangChain EmbeddingsFilter needs at least one stopping criterion - either keep top-k documents or keep those above a similarity threshold - otherwise its behavior is undefined.

Source

Thrown at packages/components/nodes/retrievers/EmbeddingsFilterRetriever/EmbeddingsFilterRetriever.ts:101

            {
                label: 'Text',
                name: 'text',
                description: 'Concatenated string from pageContent of documents',
                baseClasses: ['string', 'json']
            }
        ]
    }

    async init(nodeData: INodeData, input: string): Promise<any> {
        const baseRetriever = nodeData.inputs?.baseRetriever as BaseRetriever
        const embeddings = nodeData.inputs?.embeddings as Embeddings
        const query = nodeData.inputs?.query as string
        const similarityThreshold = nodeData.inputs?.similarityThreshold as string
        const k = nodeData.inputs?.k as string
        const output = nodeData.outputs?.output as string

        if (k === undefined && similarityThreshold === undefined) {
            throw new Error(`Must specify one of "k" or "similarity_threshold".`)
        }

        const similarityThresholdNumber = similarityThreshold ? parseFloat(similarityThreshold) : 0.8
        const kNumber = k ? parseFloat(k) : undefined

        const baseCompressor = new EmbeddingsFilter({
            embeddings: embeddings,
            similarityThreshold: similarityThresholdNumber,
            k: kNumber
        })

        const retriever = new ContextualCompressionRetriever({
            baseCompressor,
            baseRetriever: baseRetriever
        })

        if (output === 'retriever') return retriever
        else if (output === 'document') return await retriever._getRelevantDocuments(query ? query : input)

View on GitHub (pinned to abe4a8601a)

Solutions

1. Set 'k' (e.g. 4) to keep the top-k most similar documents, OR set 'similarityThreshold' (e.g. 0.8) to keep documents above that cosine threshold.
2. Set both for combined behavior, but at least one is required.
3. If driving from a variable, ensure it resolves to a non-empty string.

Example fix

// before: both undefined
// after: in node config, set k = '4' (or similarityThreshold = '0.8')

function pickStopCriterion(k?: string, threshold?: string): { k?: number; similarityThreshold?: number } {
  if (k === undefined && threshold === undefined) {
    throw new Error('Set at least one of k or similarityThreshold on the Embeddings Filter Retriever.')
  }
  return { k: k ? parseFloat(k) : undefined, similarityThreshold: threshold ? parseFloat(threshold) : undefined }
}

function hasStopCriterion(nodeData: { inputs?: Record<string, unknown> }): boolean {
  const { k, similarityThreshold } = (nodeData.inputs ?? {})
  return k !== undefined && k !== '' || similarityThreshold !== undefined && similarityThreshold !== ''
}

// Config error, not retryable. Catch only to surface a UI hint.
try {
  await retriever.init(nodeData, input)
} catch (e) {
  if (e instanceof Error && /k" or "similarity_threshold/.test(e.message)) {
    // highlight the k / similarityThreshold field
  }
  throw e
}

When it happens

Trigger: Both the 'k' and 'similarityThreshold' fields left blank on the Embeddings Filter Retriever node; one field connected to an empty upstream output; values explicitly set to undefined via API.

Common situations: New node left at defaults expecting sensible behavior; user removed the value thinking the other field would compensate; chatflow wired with a variable that resolves to empty.

Related errors

• Agent name is required!
• No extraction path configured
• JSON Path is required
• No Jira host provided
• MCP Server Config is required

AI-assisted analysis of FlowiseAI/Flowise@abe4a8601a (2026-08-12). Data as JSON: /api/errors/5428775a03a3c8f9. Report an issue: GitHub.