px
pypx

exa-py

v2.28.0

Python SDK for Exa API.

142 KB Python 3.9+ py3-none-anyMIT License7 deps
Diff
$ uv add exa-py

Exa Python SDK

PyPI version

The official Python SDK for Exa, the web search API for AI.

Documentation | Dashboard

Install

pip install exa-py

Requires Python 3.9+

Quick Start

from exa_py import Exa

exa = Exa(api_key="your-api-key")

# Search the web
results = exa.search(
    "blog post about artificial intelligence",
    type="auto",
    contents={"highlights": True}
)

# Ask a question
response = exa.answer("What is the capital of France?")

Search

results = exa.search(
    "machine learning startups",
    contents={"highlights": True}
)
results = exa.search(
    "climate tech news",
    num_results=20,
    start_published_date="2024-01-01",
    include_domains=["techcrunch.com", "wired.com"],
    contents={"highlights": True}
)
results = exa.search(
    "What are the latest battery breakthroughs?",
    type="auto",
    system_prompt="Prefer official sources and avoid duplicate results",
    output_schema={
        "type": "object",
        "properties": {
            "summary": {"type": "string"},
            "key_companies": {"type": "array", "items": {"type": "string"}},
        },
        "required": ["summary", "key_companies"],
    },
)
print(results.output.content if results.output else None)
for chunk in exa.stream_search(
    "What are the latest battery breakthroughs?",
    type="auto",
    output_schema={"type": "text"},
):
    if chunk.content:
        print(chunk.content, end="", flush=True)

Search output_schema modes:

  • {"type": "text", "description": "..."}: return plain text in output.content
  • {"type": "object", ...}: return structured JSON in output.content

system_prompt and output_schema are supported on every search type except lite. Search streaming is available via stream_search(...), which yields OpenAI-style chat completion chunks. A search without output_schema has no text to stream. Its stream yields one chunk with the results as citations.

For type: "object", search currently enforces:

  • max nesting depth: 2
  • max total properties: 10

Deep search variants that also support additional_queries:

  • deep-lite
  • deep
  • deep-reasoning

Raw body fields

Use extra_body for per-call search body overrides. Every other method takes it too; see Request options.

result = exa.search("AI research", extra_body={"futureField": True})

Contents

results = exa.get_contents(
    ["https://exa.ai/docs"],
    text=True
)
results = exa.get_contents(
    ["https://arxiv.org/abs/2303.08774"],
    highlights=True
)

Answer

response = exa.answer("What caused the 2008 financial crisis?")
print(response.answer)
stream = exa.stream_answer("Explain quantum computing")
for chunk in stream:
    print(chunk, end="", flush=True)

# The cost arrives at the end of the stream.
print(stream.cost_dollars)

Web Search and Contents tools

Use Exa as a web_search tool in an OpenAI or Anthropic loop. Call web_search() with no arguments to get Exa's recommended settings for agentic search (type="auto" and contents={"highlights": True}).

from exa_py import Exa
from openai import OpenAI

exa = Exa()
openai_client = OpenAI()

messages = [{"role": "user", "content": "What's the latest on AI chips?"}]

completion = openai_client.chat.completions.create(
    model="gpt-5.6",
    messages=messages,
    tools=[exa.openai.web_search()],
)

message = completion.choices[0].message
messages.append(message)
messages += exa.openai.handle_tool_calls(message)
import anthropic

client = anthropic.Anthropic()
response = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=1024,
    messages=messages,
    tools=[exa.anthropic.web_search()],
)

Pass name (and optionally description) to rename the tool. Anthropic requires tool names to be unique, so a custom name lets the Exa tool run alongside Anthropic's built-in web_search_20250305 tool:

response = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=1024,
    messages=messages,
    tools=[
        exa.anthropic.web_search(name="exa_web_search"),
        {"type": "web_search_20250305", "name": "web_search", "max_uses": 5},
    ],
)

get_contents is available in the same namespaces and lets the model read pages it already has URLs for. It takes a list of URLs and accepts every Exa.get_contents option:

completion = openai_client.chat.completions.create(
    model="gpt-5.6",
    messages=messages,
    tools=[
        exa.openai.web_search(),
        exa.openai.get_contents(summary=True, max_age_hours=1),
    ],
)

For the OpenAI Responses API, use exa.openai.responses.web_search() and the same handle_tool_calls helper. The handlers answer every tool call: calls naming a tool they can't resolve get an Error: unknown tool "<name>" output instead of being dropped, so follow-up requests stay valid. If you run other tools alongside Exa's, replace those error outputs with your own results before the next request.

Agent API

The Agent API is available without a beta header.

run = exa.agent.runs.create(
    query="Find engineering leaders at AI infrastructure companies that raised a Series A or B in the last 6 months.",
    output_schema={
        "type": "object",
        "properties": {
            "people": {
                "type": "array",
                "maxItems": 10,
                "items": {
                    "type": "object",
                    "properties": {
                        "name": {"type": "string"},
                        "contact_email": {"type": "string", "format": "email"},
                        "linkedin_url": {"type": "string", "format": "uri"},
                    },
                    "required": ["name", "linkedin_url"],
                },
            }
        },
        "required": ["people"],
    },
    effort="auto",
)

run = exa.agent.runs.poll_until_finished(run.id)
print(run.output.structured if run.output else None)

For Agent Ultra, set effort="ultra":

from exa_py import Exa

exa = Exa()
run = exa.agent.runs.create(
    query="Find all companies building browser automation tools in the United States.",
    effort="ultra",
    budget={"maxCostDollars": 10},
)

Agent Monitors (Beta)

Agent Monitors use the beta namespace and require the AGENT_MONITORS_BETA_HEADER beta identifier (agent-monitors-2026-08-04).

An Agent Monitor keeps a table of entities × fields fresh on a cadence.

from exa_py.agent import AGENT_MONITORS_BETA_HEADER

betas = [AGENT_MONITORS_BETA_HEADER]

# Create a monitor. Creation is async: it returns with status "creating"
# and becomes "active" once the first refresh completes.
monitor = exa.beta.agent.monitors.create(
    betas=betas,
    cadence="7d",
    entities=[
        {"name": "Acme Corp", "domain": "acme.com"},
        {"name": "Globex", "domain": "globex.com"},
    ],
    fields=[
        {"name": "funding", "description": "New funding rounds"},
        {"name": "ceo", "description": "The company's current CEO"},
    ],
    idempotency_key="my-monitor-1",  # safe retries: same key returns the same monitor
)

# Page the monitor's current entities and their contents.
for view in exa.beta.agent.monitors.entities.list_all(monitor.id, betas=betas):
    print(view.entity.name, view.contents)

# Follow the content change feed (resume later from the page's next_cursor).
changes = exa.beta.agent.monitors.changes.list(
    monitor.id,
    betas=betas,
    since="2026-01-01T00:00:00Z",
)

# One-shot backtest of a past news window — no monitor persists.
backtest = exa.beta.agent.monitors.backtests.create_and_wait(
    betas=betas,
    entities=[{"name": "Acme Corp", "domain": "acme.com"}],
    fields=[{"name": "funding", "description": "New funding rounds"}],
    start_time="2026-01-01T00:00:00Z",
    end_time="2026-01-08T00:00:00Z",
)
print(backtest.data)

# Add entities, inspect refresh progress, clean up.
exa.beta.agent.monitors.entities.add(
    monitor.id,
    betas=betas,
    entities=[{"name": "Initech", "domain": "initech.com"}],
)
current = exa.beta.agent.monitors.get(monitor.id, betas=betas)
print(current.status, current.refresh, current.usage)
exa.beta.agent.monitors.delete(monitor.id, betas=betas)

Async

from exa_py import AsyncExa

exa = AsyncExa(api_key="your-api-key")

results = await exa.search("async search example", contents={"highlights": True})

Request options

Every API method takes three keyword arguments on top of its own, on Exa and AsyncExa alike:

Argument Effect
extra_body Extra fields merged over the JSON request body
extra_headers Extra request headers
extra_query Extra query parameters
# Send a field the SDK does not type yet.
exa.get_contents("https://example.com", text=True, extra_body={"futureField": True})

exa.websets.list(
    limit=10,
    extra_query={"futureParam": "value"},
    extra_headers={"X-Request-Source": "nightly"},
)

They add to the request the method builds. On a name the method also sets, your value wins.

  • extra_body is sent with POST and PATCH requests. Other requests ignore it; use extra_query there.
  • extra_query skips a parameter set to None and sends True and False as true and false, from both clients.
  • A header the SDK sets, such as User-Agent, is replaced whatever its case. A header set to None is skipped.
  • The method you call decides whether the response streams. stream is never taken from extra_body or extra_query, and an extra Accept header neither starts a stream nor replaces the Accept of a streaming method: call the streaming method instead.
  • Helpers that make several requests (list_all, get_all, poll_until_finished, create_and_wait, wait_until_idle) apply them to each one, so create_and_wait sends extra_body with the create request and not with its polls. A cursor in extra_query picks the first page of list_all and get_all; the helper advances from there.
  • A CSV import's upload goes to a presigned URL and receives none of the three.
  • The LLM tool helpers (exa.tools, exa.openai, exa.anthropic) pass the options and timeout given when you create a tool to every request it makes, for example exa.openai.web_search(timeout=10, extra_headers={"X-Request-Source": "agent"}). run, execute and the handle_tool_calls / handle_tool_use helpers take none per call.

Timeouts and cancellation

The sync and async clients default to a 600-second network inactivity timeout (not a wall-clock deadline). Set a client-wide value or override it for a search, contents, answer, or streaming call; use None at construction to disable it. A per-call None inherits the client's value.

from exa_py import Exa

exa = Exa(api_key="your-api-key", timeout=30)
results = exa.search("latest AI developments", timeout=5)

A request that exceeds its timeout raises ExaTimeoutError, from the sync and the async client alike:

from exa_py import Exa, ExaTimeoutError

exa = Exa(api_key="your-api-key")
try:
    results = exa.search("latest AI developments", timeout=5)
except ExaTimeoutError:
    results = None

The error is also an instance of the exception the HTTP library raised, so handlers written for those keep working:

  • AsyncExa: an httpx.TimeoutException.
  • Exa: a requests.exceptions.Timeout while connecting or waiting for the response to start, and a requests.exceptions.ConnectionError when the response stalls after it started, as Requests reports it.

With AsyncExa, cancel an in-flight operation by cancelling the task awaiting it:

import asyncio
from exa_py import AsyncExa

exa = AsyncExa(api_key="your-api-key")
task = asyncio.create_task(exa.search("latest AI developments"))
task.cancel()

When stopping an async search/answer stream early, call await stream.aclose() to release its HTTP connection; cancellation while awaiting the next chunk also closes it. The sync client cannot cancel a blocking read from another thread; the timeout bounds each period of network inactivity. Both constructors and per-call overrides require a positive finite timeout when not None.

Warnings

The API can return warnings about a request, such as a field it ignores. At the time of writing, search responses carry them, including stream_search. A warning never changes the status code or the results. It is for whoever maintains the code that builds the request. Fix the request, and don't pass the warning to a model. The response keeps them in warnings.

The client also issues each one as an ExaWarning through Python's warnings module, attributed to the line of your code that made the call:

app.py:12: ExaWarning: [exa] POST /search IGNORED_FIELD at /startCrawlDate: startCrawlDate is no longer supported and is ignored. Remove it from the request.

Python's default filter shows a given warning once per calling line. Silence them, or route them to logging:

import logging
import warnings

from exa_py import ExaWarning

warnings.filterwarnings("ignore", category=ExaWarning)  # silence them
logging.captureWarnings(True)  # or send them to the "py.warnings" logger

Each ExaWarning has the attributes warning, endpoint, method and request_id. warning is the ApiWarning, with its code, path and message.

Deprecated options raise a DeprecationWarning before the request is sent, whether you set them as options or through extra_body. At the time of writing, those are start_crawl_date, end_crawl_date, and the highlights options num_sentences and highlights_per_url. Python shows a DeprecationWarning by default only for code run as __main__. Test runners such as pytest show them too.

More

See the full documentation for all features including websets, filters, and advanced options.

Details

Version
2.28.0
License
MIT License
Python
>=3.9
Maintainer
Exa AI

Release Cadence

42
releases in the past year
avg 8 days between releases

Maintainers