feat(perplexity): native content-block streaming events

[Nick Hollon (nick-hollon-lc)]

ChatPerplexity now implements a native stream_events(version="v3") path (_stream_chat_model_events / _astream_chat_model_events) that builds text and tool_call content blocks directly from Perplexity's streaming delta, rather than riding the compat bridge.

The converter is bespoke and adds no langchain-openai dependency (mirroring the groq approach). It reuses the existing _create_usage_metadata. Perplexity reports cumulative token usage per chunk, so the message total is taken from the last chunk that carries usage rather than accumulated across chunks.

Perplexity's search extras (citations, search_results, images, related_questions, videos, reasoning_steps) are surfaced on the assembled message's response_metadata. The v3 assembled message has no additional_kwargs channel, and the compat bridge drops additional_kwargs entirely — so this native path preserves search data that the bridge-based v3 path would otherwise lose. Reasoning text arrives inline (Perplexity emits <think>…</think> within content, with no separate field), so it surfaces as ordinary text.

The public stream_events contract is unchanged and the compat bridge remains the fallback. Surfacing the search extras as dedicated non-standard content blocks was intentionally deferred to keep this change low-risk; it can be revisited.

Reviewer note: unit tests patch the SDK at the chat.completions.create boundary (no network) and assert the cumulative-usage "last total wins" behavior plus the round-trip of citations/model_name onto the assembled message's response_metadata.

[open-swe]

Open SWE Review found 1 potential issue.

Open in Web • View Open SWE trace

[open-swe]

🟡 Empty-choice streams lose usage events

This continue means the native v3 path never emits message-start or message-finish if Perplexity returns only choices: [] chunks (for example an empty completion that still carries usage/metadata). Before this native hook existed, stream_events(version="v3") went through _stream, whose empty-choice branch intentionally yielded an empty chunk so the stream was not empty and downstream callers still received usage metadata. With the hook installed, the same response now produces no events and loses the final usage; the async converter has the same branch.

(Refers to lines 131-132)

---

Your feedback helps Open SWE learn. React with 👍 or 👎 to tell us if this review comment was useful.

[open-swe]

Open SWE Review found 1 potential issue.

Open in Web • View Open SWE trace

[open-swe]

🟠 V3 events bypass Responses routing

The native stream_events(version="v3") hook always calls Chat Completions here instead of applying the new _use_responses_api / _to_responses_payload routing used by _stream and _generate. With ChatPerplexity(use_responses_api=True).stream_events(..., version="v3"), or the documented auto-routing case stream_events(..., tools=[{"type": "web_search"}], version="v3"), core selects this native hook and sends Responses-only fields/built-in tools to chat.completions.create, which this PR's own routing logic says Chat Completions rejects. The async hook has the same direct Chat Completions call.

(Refers to lines 1398-1400)

---

Your feedback helps Open SWE learn. React with 👍 or 👎 to tell us if this review comment was useful.

[Mohamed-Taqy Salmi (mohamedtaqysalmi)]

The Perplexity native stream converter documents important provider quirks upfront: cumulative usage on the last chunk, inline redacted_thinking as text, and citations moved to response_metadata on the v3 path. Using BlockStreamTracker keeps this consistent with other partner streaming PRs in the series.

[Mason Daugherty (mdrxy)]

Mohamed-Taqy Salmi (@mohamedtaqysalmi) stop posting unsolicited comments, thanks