Add Perplexity Search API tool + Perplexity Embeddings API provider#1903
Add Perplexity Search API tool + Perplexity Embeddings API provider#1903jliounis wants to merge 9 commits into
Conversation
|
Note Reviews pausedIt looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the Use the following commands to manage reviews:
Use the checkboxes below for quick actions:
WalkthroughAdds a new Perplexity internet search tool: config schema, async HTTP integration with retries and backoff, package registration, docs, tests, and a Vale vocabulary entry for the token "Perplexity". (47 words) ChangesPerplexity Search Tool Integration
Sequence DiagramsequenceDiagram
participant Workflow as Workflow / User
participant Tool as Perplexity Tool
participant Config as Tool Config
participant HTTP as httpx.AsyncClient
participant API as Perplexity Search API
Workflow->>Tool: Call perplexity_internet_search(config, builder)
Tool->>Config: Resolve API key (config or env)
alt API key empty
Tool-->>Workflow: Return "unavailable" message
else
Tool->>Tool: Truncate query if needed
Tool->>Tool: Build headers (auth, X-Pplx-Integration)
loop Retry with backoff (up to max_retries)
Tool->>HTTP: POST to Perplexity Search API
HTTP->>API: Send search request
alt API success
API-->>HTTP: Return results
HTTP-->>Tool: Parse response
Tool->>Tool: Format results to <Document> blocks
Tool-->>Workflow: Return formatted output
else API error
API-->>HTTP: Error response
HTTP-->>Tool: Retry or fail
end
end
alt All retries exhausted
Tool-->>Workflow: Return failure message
end
end
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes 🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Tip 💬 Introducing Slack Agent: The best way for teams to turn conversations into code.Slack Agent is built on CodeRabbit's deep understanding of your code, so your team can collaborate across the entire SDLC without losing context.
Built for teams:
One agent for your entire SDLC. Right inside Slack. Comment |
Signed-off-by: PSI Bot <psi@perplexity.ai> Signed-off-by: James <james.liounis@perplexity.ai>
jliounis
force-pushed
the
add-perplexity-internet-search
branch
from
535ac26 to
9b8f95f
Compare
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
packages/nvidia_nat_langchain/tests/test_perplexity_internet_search.py (1)
75-77: ⚡ Quick winAlign fixture declaration with repository pytest fixture conventions.
Please define
name=on the fixture decorator and rename the fixture function to usefixture_prefix or_fixturesuffix.As per coding guidelines, "Pytest fixtures should define the name argument ... [and] function being decorated should be named using either a `fixture_` prefix or `_fixture` suffix."Suggested change
-@pytest.fixture -def tool_config(): +@pytest.fixture(name="tool_config") +def tool_config_fixture() -> PerplexityInternetSearchToolConfig: return PerplexityInternetSearchToolConfig(api_key="test-key", max_retries=2, max_query_length=50)🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/nvidia_nat_langchain/tests/test_perplexity_internet_search.py` around lines 75 - 77, Rename the pytest fixture function tool_config to follow convention (e.g., fixture_tool_config or tool_config_fixture) and update its `@pytest.fixture` decorator to include the name argument (e.g., `@pytest.fixture`(name="tool_config")), ensuring the declared fixture name remains "tool_config" while the function uses the required fixture_ prefix or _fixture suffix; keep the returned object creation (PerplexityInternetSearchToolConfig(api_key="test-key", max_retries=2, max_query_length=50)) unchanged so tests referencing the "tool_config" fixture continue to work.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In
`@packages/nvidia_nat_langchain/src/nat/plugins/langchain/tools/perplexity_internet_search.py`:
- Line 53: The country Field currently accepts any string; add schema-level
validation to enforce ISO 3166-1 alpha-2 shape by restricting values to exactly
two ASCII letters and normalising case before use: update the Field declaration
for country to include a regex/constraint for two letters (e.g. /^[A-Za-z]{2}$/)
and add a pydantic validator for the country attribute that uppercases the value
(and returns None if input is None) so invalid values are rejected at model
construction time; reference the existing country Field and its validator in the
same model/class where country: str | None = Field(...) is declared.
---
Nitpick comments:
In `@packages/nvidia_nat_langchain/tests/test_perplexity_internet_search.py`:
- Around line 75-77: Rename the pytest fixture function tool_config to follow
convention (e.g., fixture_tool_config or tool_config_fixture) and update its
`@pytest.fixture` decorator to include the name argument (e.g.,
`@pytest.fixture`(name="tool_config")), ensuring the declared fixture name remains
"tool_config" while the function uses the required fixture_ prefix or _fixture
suffix; keep the returned object creation
(PerplexityInternetSearchToolConfig(api_key="test-key", max_retries=2,
max_query_length=50)) unchanged so tests referencing the "tool_config" fixture
continue to work.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: 1ebf83fb-a343-4069-a34b-f464d7bc500e
📒 Files selected for processing (5)
ci/vale/styles/config/vocabularies/nat/accept.txtdocs/source/get-started/tutorials/add-tools-to-a-workflow.mdpackages/nvidia_nat_langchain/src/nat/plugins/langchain/tools/perplexity_internet_search.pypackages/nvidia_nat_langchain/src/nat/plugins/langchain/tools/register.pypackages/nvidia_nat_langchain/tests/test_perplexity_internet_search.py
Signed-off-by: James Liounis <james.liounis@perplexity.ai>
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
packages/nvidia_nat_langchain/src/nat/plugins/langchain/tools/perplexity_internet_search.py (1)
75-76: ⚡ Quick winAdd a public docstring to
perplexity_internet_search.Line 76 exposes a public registration function without a docstring. Add a short Google-style docstring so the public API surface is consistently documented.
Proposed fix
`@register_function`(config_type=PerplexityInternetSearchToolConfig, framework_wrappers=[LLMFrameworkEnum.LANGCHAIN]) async def perplexity_internet_search(tool_config: PerplexityInternetSearchToolConfig, builder: Builder): + """Register the `perplexity_internet_search` LangChain tool. + + Args: + tool_config (PerplexityInternetSearchToolConfig): Tool configuration. + builder (Builder): Workflow builder context. + """ import osAs per coding guidelines, "Provide Google-style docstrings for every public module, class, function and CLI command."
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/nvidia_nat_langchain/src/nat/plugins/langchain/tools/perplexity_internet_search.py` around lines 75 - 76, Add a short Google-style docstring to the public function perplexity_internet_search to document its purpose, parameters, and return value; specifically describe what the function does, the tool_config parameter (PerplexityInternetSearchToolConfig) and builder parameter (Builder), and what the registered/returned object or side-effects are, following the existing project docstring style for public APIs.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In
`@packages/nvidia_nat_langchain/src/nat/plugins/langchain/tools/perplexity_internet_search.py`:
- Around line 117-139: The retry loop in the Perplexity search code (inside the
for loop that posts to PERPLEXITY_SEARCH_URL) currently catches all exceptions
and retries, which causes unnecessary retries for non-retriable client errors;
update the exception handling in that block to inspect HTTP responses and only
retry on transient errors (HTTP 5xx and 429) or network/timeouts, while
returning early for non-retriable 4xx errors (e.g., 401, 403) with a clear error
message; use response.raise_for_status() results or the response.status_code to
branch accordingly and avoid exponential-backoff retries for deterministic
client failures, and retain logger.exception for transient failures. Also add a
Google-style docstring to the public function perplexity_internet_search
describing its purpose, parameters, and return value.
---
Nitpick comments:
In
`@packages/nvidia_nat_langchain/src/nat/plugins/langchain/tools/perplexity_internet_search.py`:
- Around line 75-76: Add a short Google-style docstring to the public function
perplexity_internet_search to document its purpose, parameters, and return
value; specifically describe what the function does, the tool_config parameter
(PerplexityInternetSearchToolConfig) and builder parameter (Builder), and what
the registered/returned object or side-effects are, following the existing
project docstring style for public APIs.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: a7b2d2b1-aa99-47e2-9249-08a4186b8b9b
📒 Files selected for processing (2)
packages/nvidia_nat_langchain/src/nat/plugins/langchain/tools/perplexity_internet_search.pypackages/nvidia_nat_langchain/tests/test_perplexity_internet_search.py
✅ Files skipped from review due to trivial changes (1)
- packages/nvidia_nat_langchain/tests/test_perplexity_internet_search.py
…tion-header helper Addresses CodeRabbit review nit: public registration function and the integration-header helper now carry Google-style docstrings, raising docstring coverage on the new module above NeMo's 80% threshold. Signed-off-by: James Liounis <james.liounis@perplexity.ai>
|
Pushed Addressed
Declined (with reason)
Still need maintainer help to unblock CI (admin-only):
DCO is green. Tests pass locally (matching the Exa pattern line-for-line). Happy to address anything else. |
There was a problem hiding this comment.
🧹 Nitpick comments (1)
packages/nvidia_nat_langchain/src/nat/plugins/langchain/tools/perplexity_internet_search.py (1)
82-82: ⚡ Quick winAdd a return type hint to the public
perplexity_internet_searchAPI.The function is publicly registered and currently omits an explicit return type annotation.
🔧 Suggested change
+from collections.abc import AsyncIterator @@ -async def perplexity_internet_search(tool_config: PerplexityInternetSearchToolConfig, builder: Builder): +async def perplexity_internet_search( + tool_config: PerplexityInternetSearchToolConfig, builder: Builder +) -> AsyncIterator[FunctionInfo]:As per coding guidelines, "All public APIs require Python 3.11+ type hints on parameters and return values."
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/nvidia_nat_langchain/src/nat/plugins/langchain/tools/perplexity_internet_search.py` at line 82, The public async function perplexity_internet_search is missing an explicit return type hint; update its signature to include the concrete return annotation that matches what it actually returns (for example -> PerplexityInternetSearchTool or -> Tool or the specific class/type returned) and import that type if necessary; if the exact return type is not yet defined, use a precise typing annotation such as -> typing.Coroutine[typing.Any, typing.Any, ReturnType] or -> typing.Any as a temporary measure, then run static type checks and replace with the concrete return type referenced in the function body (refer to the perplexity_internet_search function and its returned value to pick the correct symbol).
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Nitpick comments:
In
`@packages/nvidia_nat_langchain/src/nat/plugins/langchain/tools/perplexity_internet_search.py`:
- Line 82: The public async function perplexity_internet_search is missing an
explicit return type hint; update its signature to include the concrete return
annotation that matches what it actually returns (for example ->
PerplexityInternetSearchTool or -> Tool or the specific class/type returned) and
import that type if necessary; if the exact return type is not yet defined, use
a precise typing annotation such as -> typing.Coroutine[typing.Any, typing.Any,
ReturnType] or -> typing.Any as a temporary measure, then run static type checks
and replace with the concrete return type referenced in the function body (refer
to the perplexity_internet_search function and its returned value to pick the
correct symbol).
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: 42315761-b95a-4262-8ed6-30cc688d7949
📒 Files selected for processing (1)
packages/nvidia_nat_langchain/src/nat/plugins/langchain/tools/perplexity_internet_search.py
|
Hey @jliounis, thanks for submitting this PR. I've been in touch with Rachel, who I presume you're working with on our end. But I wanted to see if you'd be open to a quick sync to discuss a better way to package this feature. We're thinking it makes more sense for this to be a separate Perplexity-owned repo that you guys can maintain and more-quickly bring features into without being bogged down by our review cycle. Additionally, we'd be able to make the features framework-agnostic, so utilization is not just limited to langchain agents but rather all (~8) frameworks that NAT supports. Let me know if you'd be open to it! |
|
@bbednarski9 thanks for the thoughtful framing — happy to sync. Before we lock in a plan, want to push back gently on the separate-repo piece and propose an in-tree alternative that gets you the framework-agnostic outcome without the overhead of a new repo. Why I'd lean toward keeping this in-tree:
The framework-agnostic goal is fully achievable in-tree. Concretely, I'd extend this PR to:
On the velocity concern — fully understand the worry about review cycles slowing us down. If that becomes a real issue, the cleanest mitigation is co-CODEOWNERS on Working with Rachel — happy to grab time whenever works on your end. In parallel, I'll start the in-tree refactor (shared core + 8-framework registration) and push it to this branch. If after the sync you'd still strongly prefer a separate repo, we can pivot — the refactored code ports either way. |
Per maintainer @bbednarski9's request on PR NVIDIA#1903, extend the Perplexity internet search tool from LangChain-only to all 8 frameworks supported by LLMFrameworkEnum: LANGCHAIN, LLAMA_INDEX, CREWAI, SEMANTIC_KERNEL, AGNO, ADK, STRANDS, AUTOGEN. The tool implementation is HTTP-only (raw httpx calls to api.perplexity.ai/search) with no framework-specific imports, so it can live in a single framework-agnostic location and be wrapped for each target framework via NAT's existing tool_wrapper.py mechanism in each plugin package. Changes: - Move tool from packages/nvidia_nat_langchain/.../tools/ to packages/nvidia_nat_core/.../tool/ (framework-agnostic location) - Expand framework_wrappers list to include all 8 LLMFrameworkEnum values - Move tests to nvidia_nat_core's test suite - Update docs to reflect multi-framework support
Per maintainer @bbednarski9's request on PR NVIDIA#1903, extend the Perplexity internet search tool from LangChain-only to all 8 frameworks supported by LLMFrameworkEnum: LANGCHAIN, LLAMA_INDEX, CREWAI, SEMANTIC_KERNEL, AGNO, ADK, STRANDS, AUTOGEN. The tool implementation is HTTP-only (raw httpx calls to api.perplexity.ai/search) with no framework-specific imports, so it can live in a single framework-agnostic location and be wrapped for each target framework via NAT's existing tool_wrapper.py mechanism in each plugin package. Changes: - Move tool from packages/nvidia_nat_langchain/.../tools/ to packages/nvidia_nat_core/.../tool/ (framework-agnostic location) - Expand framework_wrappers list to include all 8 LLMFrameworkEnum values - Move tests to nvidia_nat_core's test suite - Update docs to reflect multi-framework support Signed-off-by: James Liounis <james.liounis@perplexity.ai>
jliounis
force-pushed
the
add-perplexity-internet-search
branch
from
efbdafd to
f3495c0
Compare
|
Pushed Changes
Net diff is small — most of the work is the file move. Easy to review with Let me know if there's a different shared-core location you'd prefer ( |
There was a problem hiding this comment.
Actionable comments posted: 2
🧹 Nitpick comments (1)
packages/nvidia_nat_core/tests/nat/tools/test_perplexity_internet_search.py (1)
85-90: ⚡ Quick winAlign fixture declaration with repository fixture naming rules.
Please set an explicit fixture name and rename the fixture function to the required pattern.
Proposed fix
-@pytest.fixture -def tool_config(): +@pytest.fixture(name="tool_config") +def tool_config_fixture(): from nat.tool.perplexity_internet_search import PerplexityInternetSearchToolConfig return PerplexityInternetSearchToolConfig(api_key="test-key", max_retries=2, max_query_length=50)As per coding guidelines, “Pytest fixtures should define the name argument... [and] function being decorated should be named using either a
fixture_prefix or_fixturesuffix.”🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@packages/nvidia_nat_core/tests/nat/tools/test_perplexity_internet_search.py` around lines 85 - 90, Rename the fixture function and set an explicit pytest fixture name: change the function currently named tool_config to follow the repository pattern (e.g., fixture_tool_config or tool_config_fixture) and add the name argument on the decorator like `@pytest.fixture`(name="tool_config") so tests using "tool_config" keep working; update the function name in the file (the function that returns PerplexityInternetSearchToolConfig) to the new prefixed/suffixed name and keep the decorator name="tool_config" to preserve the public fixture identifier.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@docs/source/get-started/tutorials/add-tools-to-a-workflow.md`:
- Line 202: Replace the prose usage of "NAT" in the sentence that currently
reads "supported by NAT" with "the toolkit" or "NeMo Agent Toolkit" while
keeping backtick-enclosed `nat` identifiers unchanged; locate the sentence
containing "perplexity_internet_search" and update "supported by NAT" →
"supported by the toolkit" (or "supported by NeMo Agent Toolkit") so that
toolkit names appear only as full names in prose and `nat` remains only in
code/package ticks.
In `@packages/nvidia_nat_core/src/nat/tool/perplexity_internet_search.py`:
- Line 94: The public registration function perplexity_internet_search is an
async generator but lacks a return type; update its signature to return
AsyncIterator[FunctionInfo] (e.g. async def
perplexity_internet_search(tool_config: PerplexityInternetSearchToolConfig,
builder: Builder) -> AsyncIterator[FunctionInfo]:) and add the necessary imports
for AsyncIterator (from typing) and FunctionInfo (where FunctionInfo is defined
in your tool/registration types) so the function signature and exports have
explicit type hints.
---
Nitpick comments:
In `@packages/nvidia_nat_core/tests/nat/tools/test_perplexity_internet_search.py`:
- Around line 85-90: Rename the fixture function and set an explicit pytest
fixture name: change the function currently named tool_config to follow the
repository pattern (e.g., fixture_tool_config or tool_config_fixture) and add
the name argument on the decorator like `@pytest.fixture`(name="tool_config") so
tests using "tool_config" keep working; update the function name in the file
(the function that returns PerplexityInternetSearchToolConfig) to the new
prefixed/suffixed name and keep the decorator name="tool_config" to preserve the
public fixture identifier.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: 2317e054-cfd0-4615-827e-cef412381415
📒 Files selected for processing (4)
docs/source/get-started/tutorials/add-tools-to-a-workflow.mdpackages/nvidia_nat_core/src/nat/tool/perplexity_internet_search.pypackages/nvidia_nat_core/src/nat/tool/register.pypackages/nvidia_nat_core/tests/nat/tools/test_perplexity_internet_search.py
There was a problem hiding this comment.
♻️ Duplicate comments (1)
docs/source/get-started/tutorials/add-tools-to-a-workflow.md (1)
202-202:⚠️ Potential issue | 🟡 Minor | ⚡ Quick winReplace "NAT" acronym with "the toolkit" in documentation prose.
Line 202 still contains "supported by NAT" which violates the coding guideline: "Never use 'NAT' or 'nat' to refer to the toolkit in documentation." Use "the toolkit" or "NeMo Agent Toolkit" instead. Keep
natonly for code/package identifiers in backticks.📝 Proposed fix
-The `perplexity_internet_search` tool ships with the core `nvidia-nat` package and is framework-agnostic — it can be used with any of the agent frameworks supported by NAT (`langchain`, `llama_index`, `crewai`, `semantic_kernel`, `agno`, `adk`, `strands`, and `autogen`). No framework-specific extra is required to install it: +The `perplexity_internet_search` tool ships with the core `nvidia-nat` package and is framework-agnostic — it can be used with any of the agent frameworks supported by the toolkit (`langchain`, `llama_index`, `crewai`, `semantic_kernel`, `agno`, `adk`, `strands`, and `autogen`). No framework-specific extra is required to install it:As per coding guidelines, "Never use 'NAT' or 'nat' to refer to the toolkit in documentation."
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@docs/source/get-started/tutorials/add-tools-to-a-workflow.md` at line 202, Replace the phrase "supported by NAT" in the sentence that mentions the `perplexity_internet_search` tool with "supported by the toolkit" (or "supported by the NeMo Agent Toolkit") so the prose no longer uses the acronym "NAT"; keep any backticked identifiers like `nvidia-nat` or `nat` unchanged in code/package contexts and only update the surrounding documentation text (the sentence that currently reads "it can be used with any of the agent frameworks supported by NAT").
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Duplicate comments:
In `@docs/source/get-started/tutorials/add-tools-to-a-workflow.md`:
- Line 202: Replace the phrase "supported by NAT" in the sentence that mentions
the `perplexity_internet_search` tool with "supported by the toolkit" (or
"supported by the NeMo Agent Toolkit") so the prose no longer uses the acronym
"NAT"; keep any backticked identifiers like `nvidia-nat` or `nat` unchanged in
code/package contexts and only update the surrounding documentation text (the
sentence that currently reads "it can be used with any of the agent frameworks
supported by NAT").
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: cb16bec1-cd99-4be4-be2b-b13d3f2abecc
📒 Files selected for processing (4)
docs/source/get-started/tutorials/add-tools-to-a-workflow.mdpackages/nvidia_nat_core/src/nat/tool/perplexity_internet_search.pypackages/nvidia_nat_core/src/nat/tool/register.pypackages/nvidia_nat_core/tests/nat/tools/test_perplexity_internet_search.py
✅ Files skipped from review due to trivial changes (1)
- packages/nvidia_nat_core/src/nat/tool/register.py
🚧 Files skipped from review as they are similar to previous changes (2)
- packages/nvidia_nat_core/src/nat/tool/perplexity_internet_search.py
- packages/nvidia_nat_core/tests/nat/tools/test_perplexity_internet_search.py
- Validate country field as ISO 3166-1 alpha-2 (min/max length 2, pattern ^[A-Z]{2}$)
- Skip retry on non-retriable 4xx (401/403/404); continue retrying 5xx and timeouts
- Replace 'NAT' acronym with 'the toolkit' in user-facing docs (per repo guideline)
- Add AsyncIterator[FunctionInfo] return type annotation to the public registration function
Signed-off-by: James Liounis <james.liounis@perplexity.ai>
|
Thanks for the thorough review, @coderabbitai! All 4 actionable comments addressed in 470d3b4:
All 14 existing tests in |
|
Tip For best results, initiate chat on the files or code changes. 🧩 Analysis chain🏁 Script executed: Length of output: 9492
The implementation looks clean and consistent with the rest of the codebase. Good luck getting the CI gates unblocked! |
Reorder the 'Alternate Method Using a Web Search Tool' section so that perplexity_internet_search is the lead example, with tavily_internet_search and exa_internet_search retained as alternatives below. perplexity_internet_search ships with nvidia-nat-core and is framework- agnostic across all 8 LLMFrameworkEnum values, making it the only single default that works regardless of which agent framework the reader picked in the earlier tutorial steps. The Tavily and Exa subsections are kept verbatim for users who prefer those backends. Signed-off-by: James Liounis <james.liounis@perplexity.ai>
|
Hi @bbednarski9 — separate from the review thread, want to flag a small follow-on I just pushed onto this PR (commit The case for Perplexity as the canonical tutorial example:
Totally happy to drop the docs commit if you'd rather merge this PR strictly as the tool addition — just let me know and I'll revert. Thanks again for the review. |
|
Hey @jliounis thanks for the heads up. Ack on your comments above. Syncing internally to make sure everything is considered per architecture. Will update shortly |
|
@bbednarski9 Looking forward to chatting. In the meantime, is there anything needed on our end to merge this? |
Add a new `perplexity` embedder provider backed by Perplexity's dedicated Embeddings API (`POST https://api.perplexity.ai/v1/embeddings`). The provider ships in the core `nvidia-nat` package and is wired into both the LangChain and LlamaIndex framework plugins, mirroring the existing `openai` / `azure_openai` / `nim` embedders. Highlights: - `PerplexityEmbedderModelConfig` supports all four documented models (`pplx-embed-v1-0.6b`, `pplx-embed-v1-4b`, plus the contextualized `pplx-embed-context-v1-*` variants), Matryoshka `dimensions`, `batch_size`, and the two on-wire `encoding_format` options (`base64_int8`, `base64_binary`). Resolves the API key from config or `PERPLEXITY_API_KEY`. - A framework-agnostic `PerplexityEmbeddings` LangChain client batches inputs (up to 512/request), decodes the base64 payload locally to a float vector, retries transient HTTP failures with backoff, and forwards an `X-Pplx-Integration: nemo-agent-toolkit/<version>` attribution header on every request. - A `PerplexityLlamaIndexEmbedding` adapter wraps the shared client to satisfy LlamaIndex's `BaseEmbedding` interface so the same NAT config can be consumed by LlamaIndex-powered retrievers/RAG flows. - Docs: adds Perplexity to the embedder provider table and configuration examples in `docs/source/build-workflows/embedders.md`, and adds Perplexity to the LangChain/LlamaIndex embedder rows in `docs/source/components/integrations/frameworks.md`. - Tests: 8 config tests + 11 LangChain client tests + 4 LlamaIndex adapter tests (23 total) covering config validation, batching, base64_int8/binary decoding, attribution header propagation, non-retriable status short-circuit, async paths, and registration with `PERPLEXITY_API_KEY` resolution. Reference: https://docs.perplexity.ai/api-reference/embeddings-post Signed-off-by: James Liounis <james.liounis@perplexity.ai>
jliounis
changed the title
|
Expanded this PR to cover both Perplexity Search and Perplexity Embeddings — pushed as What's new in this commit
Tests — 23 new tests, all green locally:
Minimal config: embedders:
perplexity_embedder:
_type: perplexity
model_name: pplx-embed-v1-0.6b
dimensions: 512 # optional Matryoshka truncationSet |
2 participants
Summary
Adds two Perplexity API integrations to the NeMo Agent Toolkit so workflows can use Perplexity for web search and for vector embeddings out of the box.
1.
perplexity_internet_search— Search API toolMirrors the Exa integration added in #1846. Backed by Perplexity's Search API (
POST /search) and framework-agnostic so it works with every framework wrapper (langchain,llama_index,crewai,semantic_kernel,agno,adk,strands,autogen).PerplexityInternetSearchToolConfigwith configurablemax_results,max_retries,max_query_length,search_recency_filter,country,max_tokens_per_page, andapi_key(via config orPERPLEXITY_API_KEY).httpx.AsyncClientand sends theX-Pplx-Integration: nemo-agent-toolkit/<version>attribution header.perplexity_internet_searchfunction for all supported framework workflows.Minimal config:
2.
perplexity— Embeddings API providerAdds a dedicated embedder provider backed by Perplexity's Embeddings API (
POST /v1/embeddings), shipped in the corenvidia-natpackage and wired into both the LangChain and LlamaIndex framework plugins (mirroringopenai/azure_openai/nim).PerplexityEmbedderModelConfigsupports all four documented models —pplx-embed-v1-0.6b,pplx-embed-v1-4b, and the document-awarepplx-embed-context-v1-0.6b/pplx-embed-context-v1-4b— plus Matryoshkadimensions,batch_size, and the two on-wireencoding_formatoptions (base64_int8,base64_binary).PerplexityEmbeddingsLangChain client batches inputs (up to 512/request), decodes the base64 payload locally to float vectors, retries transient HTTP failures with backoff, and forwards theX-Pplx-Integrationattribution header on every request.PerplexityLlamaIndexEmbeddingadapter wraps the same client to satisfy LlamaIndex'sBaseEmbeddinginterface so the same NAT config drops into LlamaIndex-powered retrievers/RAG flows.docs/source/build-workflows/embedders.md, and to the LangChain/LlamaIndex embedder rows indocs/source/components/integrations/frameworks.md.Minimal config:
Set
PERPLEXITY_API_KEYbefore using either feature:Testing
uv run --extra test pytest packages/nvidia_nat_core/tests/nat/tools/test_perplexity_internet_search.py -v— search tool tests (existing, unchanged)uv run --extra test pytest packages/nvidia_nat_core/tests/nat/embedder/test_perplexity_embedder.py -v— 8 passing (config validation, model enumeration, dimensions/batch_size bounds, encoding format choices, secret-string masking, model alias)uv run --extra test --extra langchain pytest packages/nvidia_nat_langchain/tests/test_perplexity_embeddings.py -v— 11 passing (int8/binary decode round-trip, batching, attribution header propagation, dimensions in body, non-retriable status short-circuit, async path, env-key resolution)uv run --extra test --extra langchain --extra llama-index pytest packages/nvidia_nat_llama_index/tests/test_perplexity_embedder_llama_index.py -v— 4 passing (sync + async delegation through the adapter, env-key registration)ruff check packages/nvidia_nat_core/src/nat/embedder/perplexity_embedder.py packages/nvidia_nat_langchain/src/nat/plugins/langchain/perplexity_embeddings_client.py packages/nvidia_nat_llama_index/src/nat/plugins/llama_index/perplexity_embeddings_client.pyAPI references:
Summary by CodeRabbit
New Features
_type: perplexity) available to LangChain and LlamaIndex workflows.Documentation
Tests
Style