Add Baidu Qianfan rerank retriever

[jimmyzhuu]

Summary

This PR adds a Baidu Qianfan Rerank Retriever node for reranking vector store retrieval results with Qianfan rerank models.

Changes

• Add Baidu Qianfan Rerank Retriever under Retrievers.
• Add a dedicated Baidu Qianfan API Key credential for Qianfan bearer API keys.
• Call Qianfan v2 rerank API directly at /v2/rerank.
• Use bce-reranker-base as the default rerank model.
• Preserve Flowise output modes:
  • Retriever
  • Document
  • Text
• Attach relevance_score metadata to reranked documents.
• Fall back to original documents when the Qianfan API call fails or returns invalid indexes.

Why

Flowise already supports Baidu Qianfan chat and embeddings. Adding Qianfan rerank support completes a common RAG workflow for Baidu users: retrieve candidates from a vector store, then rerank them with a dedicated cross-encoder reranker before sending context downstream.

Verification

• node packages/components/node_modules/jest/bin/jest.js --config packages/components/jest.config.js BaiduQianfanRerank.test.ts BaiduQianfanRerankRetriever.test.ts
  • 2 test suites passed
  • 13 tests passed
• git diff --check origin/main...HEAD
  • passed
• Manual smoke test against Qianfan /v2/rerank
  • returned 2 reranked documents
  • ranked the weather-related document first for a weather query
  • returned numeric relevance scores

[gemini-code-assist]

Code Review

This pull request introduces the Baidu Qianfan Rerank Retriever component, which includes credential configuration, unit tests, and an SVG icon. The retriever integrates with the Baidu Qianfan Rerank API to reorder documents based on semantic relevance. The feedback recommends validating the presence of the API key and parsing topN as an integer to prevent silent failures. Additionally, it is suggested to log errors in BaiduQianfanRerank instead of swallowing them silently, and to use handleDocumentLoaderOutput to format the retriever's output and reduce code duplication.

[gemini-code-assist]

If both qianfanApiKey and qianfanAccessKey are missing, the node will proceed with an empty API key, leading to a failed API call that silently returns the original documents. We should throw an error if no API key is provided. Additionally, topN represents a count of documents and should be parsed as an integer using parseInt(topN, 10) rather than parseFloat(topN) to avoid potential decimal values.

        const qianfanApiKey =
            getCredentialParam('qianfanApiKey', credentialData, nodeData) || getCredentialParam('qianfanAccessKey', credentialData, nodeData)
        if (!qianfanApiKey) {
            throw new Error('Baidu Qianfan API Key is missing. Please configure it in the credential.')
        }
        const k = topN ? parseInt(topN, 10) : (baseRetriever as VectorStoreRetriever).k ?? 4

References

1. When a feature requires a specific configuration (e.g., an API key for a sandboxed environment), it is preferable to throw an error if the configuration is missing rather than silently falling back to a different implementation.

[gemini-code-assist]

Swallowing all errors silently makes debugging extremely difficult if the API key is invalid, rate limits are hit, or the model name is incorrect. Consider logging the error to console.error before returning the fallback documents so that users and administrators can troubleshoot issues.

        } catch (error) {
            console.error('Baidu Qianfan Rerank Error:', error)
            return documents
        }

[gemini-code-assist]

To reduce code duplication and leverage existing utility functions, we can use handleDocumentLoaderOutput to format the retriever's output. Note: You will also need to import handleDocumentLoaderOutput from ../../../src/utils instead of handleEscapeCharacters.

        if (output === 'retriever') return retriever
        const docs = await retriever.invoke(query ? query : input)
        return handleDocumentLoaderOutput(docs, output)

[jimmyzhuu]

Resolved the actionable review items in 84891fc:

• Added an explicit credential validation error when the Qianfan API key is missing.
• Changed topN parsing to use integer parsing and added regression coverage for decimal input.
• Added tests for both cases; the targeted rerank test suite now has 13 passing tests.

For the other two suggestions:

• The compressor still falls back to the original documents on rerank API errors without logging, matching the existing Jina/Cohere/Voyage rerank retriever behavior in this package.
• I did not switch to handleDocumentLoaderOutput because this is a retriever node, not a document loader. Existing rerank retrievers format document and text outputs directly, so this keeps the node consistent with the local retriever patterns.