GitHub Repo
LangChain Fixes Lost Perplexity Search Sources Across Four Agent API Call Paths
On October 21, LangChain released langchain-perplexity 1.4.2 to fix search sources being dropped from Agent API responses. The update covers synchronous, asynchronous, and streaming calls, restoring data used by citation interfaces and verification workflows.
LangChain released langchain-perplexity==1.4.2 on October 21, 2026, including a fix to preserve search results from the Perplexity Agent API. Previously, the integration package could return answer text normally while dropping sources supplied by the service, leaving additional_kwargs["search_results"] empty even when it should contain data. The fix is listed in the official release notes. Release notes
The issue was in the API response conversion layer. The fix description says ChatPerplexity did not fully collect search sources from Agent API output items and streaming events, affecting four call paths: invoke, ainvoke, stream, and astream. The submitter said a comparison of real API requests between 1.4.1 and the patched version confirmed that the old version omitted sources. The fix also adds regression tests for multiple batches of sources and to prevent duplicate accumulation. The related pull request has been merged. Fix and verification record
For search agents, this change preserves the machine-readable evidence behind an answer. The official integration documentation places search results in the field above and explains that requests using the built-in search tool or certain Responses parameters may automatically route to the Agent API. Engineering teams should therefore check which endpoint a request actually uses, even if they have not explicitly enabled a switching flag. As an architectural inference, lost sources could leave citation lists, audit records, or downstream URL checks without input, even when the answer displayed to users appears complete. Integration documentation
However, an empty source list does not necessarily indicate an error. Perplexity’s documentation says web search in the Agent API must be enabled through a tool or default configuration; a request containing only a model and input does not automatically retrieve web citations. After upgrading, use cases that actually enable search to check standard and streaming responses, source order, and duplicate URLs. Successful source propagation does not mean the content has been fact-checked. Agent API documentation