langchain - 💡(How to fix) Fix perplexity: Add use_responses_api flag to ChatPerplexity for Agent API support [1 pull requests]

Privileged issue

• I'm opening a feature request that already has design + working code in flight (https://github.com/langchain-ai/langchain/pull/37359).

Feature request

Add a use_responses_api: bool | None = None flag to ChatPerplexity in libs/partners/perplexity/ so requests can be routed through Perplexity's Agent API (the Perplexity-flavored Responses API) in addition to the existing Chat Completions endpoint. Mirrors ChatOpenAI.use_responses_api.

Motivation

Perplexity exposes two HTTP surfaces from the same SDK client object:

• client.chat.completions.create() → Chat Completions (today's ChatPerplexity path)
• client.responses.create() → Agent API (POST /v1/agent, alias POST /v1/responses)

The Agent API supports built-in tools (web_search, fetch_url, finance_search, people_search), instructions, input, previous_response_id, and include — none of which exist on Chat Completions. Today ChatPerplexity only calls Chat Completions, so users who want the Agent API have to drop down to the raw SDK. This is the same situation ChatOpenAI solved with its use_responses_api flag.

Proposal

• New optional field use_responses_api: bool | None = None on ChatPerplexity, behavior mirroring ChatOpenAI:
  • True → always route through client.responses.create()
  • False → always route through client.chat.completions.create()
  • None (default) → auto-detect based on payload: True if the request uses any built-in tool (any tools[*] whose type is not "function") or any of previous_response_id, instructions, input, include; False otherwise.
• Backwards-compatible: every existing ChatPerplexity(model="sonar", ...) call keeps hitting client.chat.completions.create() unchanged.
• No new class. Single ChatPerplexity keeps parity with the single-class ChatOpenAI model that LangChain users already understand.

Why a flag, not a new class

ChatPerplexityAgent would duplicate ~90% of ChatPerplexity for no win, and would break parity with how the OpenAI integration handles the same two-surfaces-one-SDK situation.

System Info

N/A — net new feature.