Improvements to `ChatPerplexity` Integration #30802

jamesliounis · 2025-04-12T17:54:30Z

TL;DR:

Add enable_search_classifier parameter
Update model default to sonar-pro
Modify example notebook to reflect best practices
Updates the Perplexity API wrapper (ChatPerplexity) by adding a new parameter, enable_search_classifier, with a default value of True. This enables the use of a classifier to determine if search is necessary for a given API call. Additionally, the default model name has been updated from the deprecated "llama-3.1-sonar-small-128k-online" to "sonar-pro", keeping the integration current with the Perplexity API. The change allows users to both modify this new flag and rely on sensible defaults while preserving the flexibility of passing any additional parameters via model_kwargs.
Updates example notebook with best practices around using Perplexity and correct links to documentation.

Dependencies:
No additional dependencies are introduced with this change.

Twitter handle:
(@LiounisJames, @PPLXDevs)

Additional steps performed:
• Ran make format, make lint, and make test to ensure that the changes adhere to the code quality and testing guidelines.

Thank you for reviewing this contribution!

vercel · 2025-04-12T17:54:35Z

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
langchain	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Apr 22, 2025 7:56pm

ccurme

Thank you!

Are we still intending to add the tools to langchain-community? Should we remove those?

jamesliounis · 2025-04-13T16:58:39Z

@ccurme Removed the tool integration. Thanks for pointing this out.

…entation

james-pplx · 2025-04-21T16:12:33Z

@ccurme anything else I can do to help with this?

ccurme · 2025-04-21T21:37:22Z

libs/partners/perplexity/langchain_perplexity/chat_models.py

@@ -168,6 +168,8 @@ class StructuredOutput(BaseModel):
    """Whether to stream the results or not."""
    max_tokens: Optional[int] = None
    """Maximum number of tokens to generate."""
+    enable_search_classifier: bool = True


When I run this I get

TypeError: Completions.create() got an unexpected keyword argument 'enable_search_classifier'

Does this need to be passed in through extra_body, as with other Perplexity-specific params not supported by the OpenAI SDK? (see examples in docs).

That's correct, it needs to be passed in the extra body. Forgot to push the commit. Fixing.

addressed these comments in this commit. I removed the enable_search_classifier param, it's better to add it to the Perplexity documentation directly and let users pass it themselves if needed.

ccurme · 2025-04-21T21:38:32Z

libs/partners/perplexity/langchain_perplexity/chat_models.py

@@ -147,9 +147,9 @@ class StructuredOutput(BaseModel):
    """  # noqa: E501

    client: Any = None  #: :meta private:
-    model: str = "llama-3.1-sonar-small-128k-online"
+    model: str = "sonar-pro"


It's not ideal to change default models and temperatures, as it changes behavior (and costs?) out from under users without warning.

Ideally we'd actually remove these defaults and just rely on whatever defaults Perplexity uses server-side. I'd be supportive of doing this in a v0.2 release.

I changed the default model because the one that was there (llama-3.1-sonar-small-128k-online) has been deprecated, we do not support it anymore. sonar-pro is our best performing model for a regular web search so it will most likely result in higher quality outputs, but if cost is a concern then we can use our cheapest alternative which is sonar. These are all of our currently supported models.

I changed the inference parameters because the ones that I added are our defaults. If we remove them altogether this would then use what we have server-side which is also ok (exactly the same thing). But specifying a model is required. Additional info here in our API ref.

IMO it's preferable to have the following:

model has no default, users must pass it in on init

temperature defaults to null and is only included in request payloads if it is explicitly specified.

That is, we mirror usage of the API.

I don't feel strongly about this if you'd prefer to maintain defaults in the LangChain integration. In either case, if we are changing the defaults we probably should bump the minor version from 0.1 to 0.2, though.

ccurme

Would like to get this in but there are still some issues:

Code does not run. Please try running tests locally:
- make test will run unit tests
- PPLX_API_KEY=pplx-... make integration_test will run integration tests (with real HTTP requests)
We are (inadvertently?) undoing multiple PRs here (see comments)
- Run make format and make lint locally to run the auto-formatter and linter, which will catch some of the errors.
Docs need to be updated to reflect most recent changes.

ccurme · 2025-04-22T17:58:58Z

libs/partners/perplexity/langchain_perplexity/chat_models.py

+        additional_kwargs = {"citations": response.citations}
+        for attr in ["images", "related_questions"]:


Why are we un-doing #30805?

ccurme · 2025-04-22T18:00:58Z

libs/partners/perplexity/langchain_perplexity/chat_models.py

@@ -39,8 +50,8 @@
 from typing_extensions import Self

 _BM = TypeVar("_BM", bound=BaseModel)
-_DictOrPydanticClass = Union[dict[str, Any], type[_BM], type]
-_DictOrPydantic = Union[dict, _BM]
+_DictOrPydanticClass = Union[Dict[str, Any], Type[_BM], Type]


Why are we un-doing #30781?

ccurme · 2025-04-22T18:05:27Z

libs/partners/perplexity/langchain_perplexity/chat_models.py

@@ -147,9 +147,9 @@ class StructuredOutput(BaseModel):
    """  # noqa: E501

    client: Any = None  #: :meta private:
-    model: str = "llama-3.1-sonar-small-128k-online"
+    model: str = "sonar-pro"


IMO it's preferable to have the following:

model has no default, users must pass it in on init

temperature defaults to null and is only included in request payloads if it is explicitly specified.

That is, we mirror usage of the API.

I don't feel strongly about this if you'd prefer to maintain defaults in the LangChain integration. In either case, if we are changing the defaults we probably should bump the minor version from 0.1 to 0.2, though.

ccurme · 2025-04-22T18:06:27Z

libs/partners/perplexity/langchain_perplexity/chat_models.py

        """Get the default parameters for calling PerplexityChat API."""
        return {
            "max_tokens": self.max_tokens,
            "stream": self.streaming,
            "temperature": self.temperature,
+            "enable_search_classifier": self.enable_search_classifier,


I believe this is now broken.

Fixed. Removed it.

ccurme · 2025-04-22T18:07:08Z

docs/docs/integrations/chat/perplexity.ipynb

+    "    decide whether to execute a full web search or bypass it, depending on the query.\n",
+    "    This helps to reduce unnecessary searches while still delivering accurate responses.\n",
+    "    \"\"\"\n",
+    "    llm = ChatPerplexity(model=\"sonar\", enable_search_classifier=True)\n",


This appears out of date.

feat(chat): bump ChatPerplexity to v0.2, require model on init, conditional params - Remove default for `model` and add a pre-init validator that errors (with model‑cards URL) if it’s not provided - Make `temperature`, `max_tokens`, `stream`, and `timeout` only appear in the request payload when explicitly set - Bump class docstring and version reference to v0.2

chore(chat_models): restore `_DictOrPydantic` alias and fix citations handling in `_generate` - Re‑introduce the `_DictOrPydantic` type alias alongside `_DictOrPydanticClass` - In `_generate`, initialize `additional_kwargs` with `{"citations": response.citations}` and always append `images` and `related_questions`

jamesliounis · 2025-04-22T19:37:41Z

I restored the undone PRs (which occurred because branch is now outdated) and have ran all the tests.

…entation

ccurme

Still seeing tests failing on CI, undoing previous PRs and other unrelated changes.

ccurme · 2025-04-28T12:54:15Z

libs/partners/perplexity/langchain_perplexity/chat_models.py

-
-        See full list of supported init args and their descriptions in the params section.
-
-        Instantiate:


Why delete these? These are standard components of our API references.

James Liounis added 3 commits April 5, 2025 20:40

deploy pplx wrapper to langchain

9c7805b

deploy pplx tool

0a51959

add search classifier, make temperature 0.2, change model

c9003fb

dosubot bot added size:L This PR changes 100-499 lines, ignoring generated files. community Related to langchain-community labels Apr 12, 2025

add best practices to nb

2f56bc5

jamesliounis changed the title ~~community: add enable_search_classifier parameter and update model default to sonar-pro~~ Add enable_search_classifier parameter | update model default to sonar-pro | modify example notebook to reflect best practices Apr 12, 2025

vercel bot deployed to Preview April 12, 2025 19:07 View deployment

add example tool calling turning off search component

126e56d

jamesliounis mentioned this pull request Apr 12, 2025

Build PPLX Search Tool integration #30692

Closed

jamesliounis changed the title ~~Perplexity Integration: Add enable_search_classifier parameter | update model default to sonar-pro | modify example notebook to reflect best practices~~ Improvements to ChatPerplexity Integration Apr 12, 2025

vercel bot deployed to Preview April 12, 2025 19:23 View deployment

ccurme reviewed Apr 13, 2025

View reviewed changes

ccurme self-assigned this Apr 13, 2025

James Liounis added 2 commits April 13, 2025 12:56

remove tool implementation

6ec00e1

Remove perplexity.py

b839db2

jamesliounis requested a review from ccurme April 13, 2025 16:58

vercel bot deployed to Preview April 13, 2025 17:16 View deployment

Merge branch 'master' into jamesliounis/update-chat-perplexity-implem…

eb61440

…entation

vercel bot deployed to Preview April 17, 2025 18:29 View deployment

ccurme reviewed Apr 21, 2025

View reviewed changes

sonar-pro --> sonar + move enable_search_classifier to extra body

bb659f3

jamesliounis requested a review from ccurme April 22, 2025 00:00

vercel bot deployed to Preview April 22, 2025 00:05 View deployment

ccurme reviewed Apr 22, 2025

View reviewed changes

remove enable_search_classifier from the response

7c3ae62

vercel bot deployed to Preview April 22, 2025 18:43 View deployment

dosubot bot added size:XL This PR changes 500-999 lines, ignoring generated files. and removed size:L This PR changes 100-499 lines, ignoring generated files. labels Apr 22, 2025

vercel bot deployed to Preview April 22, 2025 18:59 View deployment

vercel bot deployed to Preview April 22, 2025 19:13 View deployment

jamesliounis added 4 commits April 22, 2025 15:27

Update README with supported model

f8a16eb

Update perplexity.ipynb

4521103

add enable_search_classifier example

4105b2b

fix invalid json

d32d55f

Merge branch 'master' into jamesliounis/update-chat-perplexity-implem…

14200e0

…entation

vercel bot deployed to Preview April 22, 2025 19:56 View deployment

ccurme reviewed Apr 28, 2025

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Improvements to `ChatPerplexity` Integration #30802

Improvements to `ChatPerplexity` Integration #30802

jamesliounis commented Apr 12, 2025 •

edited

Loading

vercel bot commented Apr 12, 2025 •

edited

Loading

ccurme left a comment

jamesliounis commented Apr 13, 2025

james-pplx commented Apr 21, 2025

ccurme Apr 21, 2025

james-pplx Apr 21, 2025

jamesliounis Apr 21, 2025

ccurme Apr 21, 2025

james-pplx Apr 21, 2025 •

edited

Loading

ccurme Apr 22, 2025

ccurme left a comment

ccurme Apr 22, 2025

ccurme Apr 22, 2025

ccurme Apr 22, 2025

ccurme Apr 22, 2025

jamesliounis Apr 22, 2025

ccurme Apr 22, 2025

jamesliounis commented Apr 22, 2025

ccurme left a comment

ccurme Apr 28, 2025

		additional_kwargs = {"citations": response.citations}
		for attr in ["images", "related_questions"]:


		See full list of supported init args and their descriptions in the params section.

		Instantiate:

Improvements to ChatPerplexity Integration #30802

Are you sure you want to change the base?

Improvements to ChatPerplexity Integration #30802

Conversation

jamesliounis commented Apr 12, 2025 • edited Loading

vercel bot commented Apr 12, 2025 • edited Loading

ccurme left a comment

Choose a reason for hiding this comment

jamesliounis commented Apr 13, 2025

james-pplx commented Apr 21, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

james-pplx Apr 21, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ccurme left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jamesliounis commented Apr 22, 2025

ccurme left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Improvements to `ChatPerplexity` Integration #30802

Improvements to `ChatPerplexity` Integration #30802

jamesliounis commented Apr 12, 2025 •

edited

Loading

vercel bot commented Apr 12, 2025 •

edited

Loading

james-pplx Apr 21, 2025 •

edited

Loading