Skip to content

Exa API Reference (2026)

Source: docs.exa.ai → redirects to exa.ai/docs LLM Discovery: exa.ai/docs/llms.txt (150+ pages) OpenAPI Specs: - Search API: exa-openapi-spec.yaml - Websets API: exa-websets-spec.yaml Python SDK: pip install exa-py (module: exa_py) Last Verified: 2026-01-13 Verified Against: Official docs via llms.txt, Exa OpenAPI specs, and live API probes where noted

Overview

Exa is "a search engine made for AIs" - optimized for RAG, agentic workflows, and structured data extraction. Unlike traditional search engines (optimized for human clicks), Exa is optimized for machine consumption.

Core Capabilities:

  • Search: Neural/embeddings-based web search with 4 modes (auto, neural, fast, deep)
  • Contents: Clean, parsed text from URLs with live crawling
  • Find Similar: Discover semantically related pages
  • Answer: LLM-generated answers with citations
  • Research: Async deep research with structured output (agentic)

Base URL & Authentication

Base URL: https://api.exa.ai

Authentication: 

x-api-key: YOUR_EXA_API_KEY

Or use Bearer token: 

Authorization: Bearer YOUR_EXA_API_KEY

Search Endpoint

POST /search

Intelligently find webpages using embeddings-based search.

Search Types

Type Description
auto Default. Intelligently selects the best method
neural Embeddings-based semantic search
fast Streamlined search optimized for speed
deep Multi-pass with query expansion for highest quality

When to use deep: Complex research queries where quality matters more than speed. The endpoint agentically searches, processes, and searches again until it finds the highest quality information.

Request Body

Parameter Type Default Description
query string Required The search query
type enum "auto" auto, neural, fast, deep
additionalQueries string[] - Extra query variations (deep search only)
numResults integer 10 Max 100 results
includeDomains string[] - Filter to specific domains
excludeDomains string[] - Exclude domains
startPublishedDate ISO 8601 - Filter by publish date
endPublishedDate ISO 8601 - Filter by publish date
startCrawlDate ISO 8601 - Filter by crawl date
endCrawlDate ISO 8601 - Filter by crawl date
includeText string[] - Must contain (1 string, 5 words max)
excludeText string[] - Must not contain (1 string, 5 words max; checks first 1,000 words)
category enum - See category table below
userLocation string - Two-letter ISO country code
moderation boolean false Moderate results for safety (SDK)
flags string[] - Experimental flags (SDK)
context boolean/object - Return combined context string for RAG
contents object - Control text/highlights/summary retrieval

Categories

Category Description
news News articles
research paper Academic papers
pdf PDF documents
github GitHub repositories
tweet Twitter/X posts
personal site Personal websites/blogs
financial report Financial documents
company Company profiles
people People profiles

Notes: The official exa-py SDK exposes moderation and flags, but they are not currently documented in exa-openapi-spec.yaml v1.2.0. Prefer treating them as optional/experimental until confirmed in vendor docs.

Response

{
  "requestId": "string",
  "results": [
    {
      "id": "string",
      "title": "string",
      "url": "string",
      "publishedDate": "string|null",
      "author": "string|null",
      "score": 0.0,
      "image": "string|null",
      "favicon": "string|null",
      "text": "string",
      "summary": "string",
      "highlights": ["string"],
      "highlightScores": [0.0]
    }
  ],
  "context": "string (combined content for RAG, if requested)",
  "searchType": "string",
  "costDollars": {
    "total": 0.005,
    "breakDown": [
      {
        "search": 0.005,
        "contents": 0.0,
        "breakdown": {
          "neuralSearch": 0.005,
          "deepSearch": 0.0,
          "contentText": 0.0,
          "contentHighlight": 0.0,
          "contentSummary": 0.0
        }
      }
    ],
    "perRequestPrices": {
      "neuralSearch_1_25_results": 0.005,
      "neuralSearch_26_100_results": 0.025,
      "deepSearch_1_25_results": 0.015,
      "deepSearch_26_100_results": 0.075
    },
    "perPagePrices": {
      "contentText": 0.001,
      "contentHighlight": 0.001,
      "contentSummary": 0.001
    }
  }
}

Note: searchType indicates which search method was used. For type="auto", this shows the actual method selected (e.g., "neural" or "deep").

Contents note: You may see text: true used at the top level in examples; the official SDKs send content options under contents (e.g. contents: {"text": true}), which supports advanced options like highlights, summary, subpages, and extras.

Code Examples

cURL: 

curl -X POST 'https://api.exa.ai/search' \
  -H 'x-api-key: YOUR_EXA_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{"query": "Latest research in LLMs", "text": true}'

Python: 

from exa_py import Exa

exa = Exa('YOUR_EXA_API_KEY')
results = exa.search("Latest research in LLMs", num_results=10, contents={"text": True})

for result in results.results:
    print(f"{result.title}: {result.url}")

JavaScript: 

import Exa from 'exa-js';
const exa = new Exa('YOUR_EXA_API_KEY');

const results = await exa.searchAndContents('Latest research in LLMs', { text: true });

Contents Endpoint

POST /contents

Obtain clean, parsed content from URLs with automatic live crawling fallback.

Livecrawl Options

Option Description
never Disable live crawling (use cache only)
fallback Livecrawl only when cache is empty
preferred Recommended. Try livecrawl first, fall back to cache if crawling fails
always Always live-crawl (not recommended without consulting Exa)
auto Let Exa choose behavior (SDK)

Request Body

Parameter Type Description
urls string[] Required. URLs to crawl
ids string[] Deprecated. Use urls instead
text boolean/object Full page text. Object: {maxCharacters, includeHtmlTags}
highlights object Extract snippets: {query, numSentences, highlightsPerUrl}
summary object LLM summaries: {query, schema} for structured output
metadata boolean/object Request metadata (SDK)
livecrawl enum never, fallback, preferred, always, auto
livecrawlTimeout integer Milliseconds (default 10000)
filterEmptyResults boolean Drop empty results (SDK)
subpages integer Number of subpages to crawl
subpageTarget string/string[] Keywords for subpage filtering
extras object Additional data: {links: N, imageLinks: N}
flags string[] Experimental flags (SDK)
context boolean/object Return combined context string for RAG

Response

{
  "requestId": "string",
  "results": [
    {
      "id": "string",
      "title": "string",
      "url": "string",
      "publishedDate": "string|null",
      "author": "string|null",
      "score": 0.0,
      "image": "string|null",
      "favicon": "string|null",
      "text": "string",
      "highlights": ["string"],
      "highlightScores": [0.0],
      "summary": "string",
      "subpages": [],
      "extras": {"links": [], "imageLinks": []}
    }
  ],
  "statuses": [
    {
      "id": "string",
      "status": "success|error",
      "error": {
        "tag": "CRAWL_NOT_FOUND|CRAWL_TIMEOUT|CRAWL_LIVECRAWL_TIMEOUT|SOURCE_NOT_AVAILABLE|CRAWL_UNKNOWN_ERROR",
        "httpStatusCode": 404
      }
    }
  ],
  "context": "string (combined content for RAG, if requested)",
  "costDollars": {
    "total": 0.001,
    "perPagePrices": {
      "contentText": 0.001,
      "contentHighlight": 0.001,
      "contentSummary": 0.001
    }
  }
}

Code Examples

Python: 

from exa_py import Exa

exa = Exa('YOUR_EXA_API_KEY')
results = exa.get_contents(
    urls=["https://arxiv.org/abs/2307.06435"],
    text=True
)

Find Similar Endpoint

POST /findSimilar

Find semantically related pages to a given URL. Supports the same filtering options as Search.

Request Body

Parameter Type Description
url string Required. Source URL to find similar pages for
numResults integer Max 100, default 10
includeDomains string[] Limit to domains
excludeDomains string[] Exclude domains
startPublishedDate ISO 8601 Filter by publish date
endPublishedDate ISO 8601 Filter by publish date
startCrawlDate ISO 8601 Filter by crawl date
endCrawlDate ISO 8601 Filter by crawl date
includeText string[] Must contain (1 string, 5 words max)
excludeText string[] Must not contain (1 string, 5 words max; checks first 1,000 words)
excludeSourceDomain boolean Exclude results from the source URL's domain (SDK)
category enum A data category to focus on (SDK)
flags string[] Experimental flags (SDK)
context boolean/object Return combined context string for RAG
contents object Configure text/highlights/summary retrieval

Response

{
  "requestId": "string",
  "results": [
    {
      "id": "string",
      "title": "string",
      "url": "string",
      "publishedDate": "string|null",
      "author": "string|null",
      "score": 0.0,
      "image": "string|null",
      "favicon": "string|null",
      "text": "string",
      "summary": "string",
      "highlights": ["string"],
      "highlightScores": [0.0]
    }
  ],
  "context": "string (combined content for RAG, if requested)",
  "costDollars": {
    "total": 0.005
  }
}

Code Examples

Python: 

from exa_py import Exa

exa = Exa('YOUR_EXA_API_KEY')
results = exa.find_similar(
    "https://arxiv.org/abs/2307.06435",
    num_results=10,
    contents={"text": True},
)

Answer Endpoint

POST /answer

Generate LLM-powered answers with citations from web search.

Request Body

Parameter Type Default Description
query string Required The question to answer
stream boolean false Server-sent events stream
text boolean false Include full text in citations

Response

{
  "answer": "string - generated answer",
  "citations": [
    {
      "id": "string",
      "url": "string",
      "title": "string",
      "author": "string|null",
      "publishedDate": "string|null",
      "image": "string|null",
      "favicon": "string|null",
      "text": "string"
    }
  ],
  "costDollars": {
    "total": 0.005,
    "breakDown": [
      {
        "search": 0.005,
        "contents": 0.0,
        "breakdown": {
          "neuralSearch": 0.005,
          "deepSearch": 0.0,
          "contentText": 0.0,
          "contentHighlight": 0.0,
          "contentSummary": 0.0
        }
      }
    ]
  }
}

Code Examples

Python: 

from exa_py import Exa

exa = Exa('YOUR_EXA_API_KEY')
result = exa.answer(
    "What is the latest valuation of SpaceX?",
    text=True
)

print(result.answer)
for citation in result.citations:
    print(f"  - {citation.title}: {citation.url}")

Research Endpoints

  • GET /research/v1 (list)
  • POST /research/v1 (create)
  • GET /research/v1/{researchId} (get / stream)

Async deep research with structured output support.

List Parameters (GET /research/v1)

Parameter Type Default Description
cursor string - Cursor for pagination (use nextCursor from prior response)
limit int 10 Number of results (OpenAPI: 1–50)

List Response (200)

{
  "data": [],
  "hasMore": false,
  "nextCursor": null
}

Note: The list endpoint wraps items under data (not items) and uses nextCursor (not cursor).

Create Request Body

Parameter Type Description
instructions string Required. Research guidelines (max 4096 chars)
model enum exa-research-fast, exa-research, exa-research-pro (official docs list all three; the Exa OpenAPI spec repo currently omits exa-research-fast, but the API accepts it — verified via live /research/v1 create on 2026-01-13)
outputSchema object JSON Schema for structured output

Get Parameters

Parameter Type Default Description
stream boolean false Stream Server-Sent Events (SSE)
events boolean false Include events in non-streaming responses

OpenAPI note: The raw exa-openapi-spec.yaml currently marks stream/events as required query params and types them as string; observed API usage (and SDKs) treat them as optional booleans.

Response (create: 201)

{
  "researchId": "string",
  "status": "pending|running|completed|canceled|failed",
  "createdAt": 1234567890123,
  "model": "exa-research",
  "instructions": "string"
}

When completed: 

{
  "researchId": "string",
  "status": "completed",
  "output": {
    "content": "string - research results",
    "parsed": {}
  },
  "costDollars": {
    "total": 0.10,
    "numSearches": 6,
    "numPages": 20,
    "reasoningTokens": 1000
  }
}

Note: When you provide outputSchema, the response may include a structured output.parsed in addition to output.content.

Code Examples

Python: 

from exa_py import Exa

exa = Exa('YOUR_EXA_API_KEY')

# Create research task
research = exa.research.create(
    instructions="Summarize latest AI safety research",
    model="exa-research"
)

# Poll for results
result = exa.research.poll_until_finished(research.research_id)

Python SDK Reference

Installation

pip install exa-py
# or
uv add exa-py

Quick Start

from exa_py import Exa

# Initialize client
exa = Exa('YOUR_EXA_API_KEY')

# Search (returns text contents by default)
results = exa.search("prediction markets research", num_results=10)

# Access results
for r in results.results:
    print(f"Title: {r.title}")
    print(f"URL: {r.url}")
    print(f"Text: {r.text[:200]}...")
    print()

Available Methods

Method Description
search(query, **kwargs) Search (API returns content fields only when requested via text/contents)
search_and_contents(query, **kwargs) Search + contents convenience method (used in OpenAPI examples v1.2.0)
get_contents(urls, **kwargs) Get contents from URLs
find_similar(url, **kwargs) Find similar pages
find_similar_and_contents(url, **kwargs) Find similar + contents convenience method (used in OpenAPI examples v1.2.0)
answer(query, **kwargs) Get LLM answer with citations
stream_answer(query, **kwargs) Streaming answer (yields chunks)
research.create(instructions=..., **kwargs) Start async research
research.get(research_id, **kwargs) Get research results
research.list(cursor=..., limit=...) List research tasks
research.poll_until_finished(research_id, **kwargs) Poll for completion

Note: Structured summaries using summary={"schema": {...}} return JSON strings that require parsing.

Pricing (as of 2026)

See Exa Pricing for current free credits and plan limits.

Search Pricing (per 1,000 requests)

Search Type 1-25 results 26-100 results
Fast/Auto/Neural $5 $25
Deep $15 $75

Contents Pricing (per 1,000 pages)

Content Type Cost
Text $1
Highlights $1
Summary $1

Answer Pricing

The Answer endpoint returns costDollars in the response. Use costDollars.total as the source of truth.

Research API Pricing (Variable)

Research API uses consumption-based billing. You're only charged for tasks that complete successfully.

Component exa-research exa-research-pro
Agent searches (per 1k) $5 $5
Agent page reads (per 1k pages*) $5 $10
Reasoning tokens (per 1M) $5 $5

*Page = 1,000 tokens of webpage content

Typical task cost: $0.10-$0.50 depending on complexity (20-40 second completion)

Rate Limits

Endpoint Limit
/search 5 QPS
/contents 50 QPS
/answer 5 QPS
/findSimilar 5 QPS (assumed, not documented)
/research/v1 15 concurrent tasks

Note: QPS = Queries Per Second. Research API uses concurrent task limits for long-running operations. Contact hello@exa.ai for Enterprise rate limit increases.

Error Codes

HTTP Status Errors

Code Meaning Action
400 Invalid request parameters, malformed JSON Validate request format
401 Missing or invalid API key Verify credentials
403 Valid key but insufficient permissions or rate exceeded Check account/throttle
404 Resource not found Confirm resource exists
409 Resource already exists (Websets) Use different identifier
429 Rate limit exceeded Implement exponential backoff
500 Server error Retry after delay
502 Upstream server issue Retry after delay
503 Service temporarily down Wait and retry

Contents Endpoint Status Tags

Errors in /contents appear in the statuses field:

Tag HTTP Code Meaning
CRAWL_NOT_FOUND 404 URL content unavailable
CRAWL_TIMEOUT 408 Fetch operation timed out
CRAWL_LIVECRAWL_TIMEOUT 408 Live crawl timed out
SOURCE_NOT_AVAILABLE 403 Access denied or paywalled
CRAWL_UNKNOWN_ERROR 500+ Other crawling failures

Tool Use with Claude (Manual)

For custom tool definitions (or when you need fine-grained control):

import anthropic
from exa_py import Exa

exa = Exa('YOUR_EXA_API_KEY')
client = anthropic.Anthropic()

# Define Exa as a tool
tools = [
    {
        "name": "web_search",
        "description": "Search the web for current information using Exa",
        "input_schema": {
            "type": "object",
            "properties": {
                "query": {
                    "type": "string",
                    "description": "The search query"
                }
            },
            "required": ["query"]
        }
    }
]

# In your tool execution loop
def execute_tool(tool_name, tool_input):
    if tool_name == "web_search":
        results = exa.search(
            tool_input["query"],
            num_results=5,
            contents={"text": True},
        )
        return "\n\n".join([
            f"**{r.title}**\n{r.url}\n{r.text[:500]}"
            for r in results.results
        ])

Integration Tips for Kalshi Research

Event Resolution

When checking if a market resolved, use category="news" and date filters:

# Find resolution sources for a market expiring this week
results = exa.search(
    "Federal Reserve interest rate decision January 2026",
    type="deep",  # Quality matters for resolution
    category="news",
    start_published_date="2026-01-06T00:00:00Z",
    include_domains=["federalreserve.gov", "reuters.com", "bloomberg.com"],
    contents={"text": True},
)

Thesis Research

For creating trading theses, use the Research API with structured output:

research = exa.research.create(
    instructions="""
    Analyze the likelihood of [EVENT] occurring by [DATE].

    Research:
    1. Historical precedent
    2. Current indicators
    3. Expert opinions
    4. Contrarian arguments

    Provide a probability estimate with confidence interval.
    """,
    model="exa-research",
    output_schema={
        "type": "object",
        "properties": {
            "probability": {"type": "number"},
            "confidence_low": {"type": "number"},
            "confidence_high": {"type": "number"},
            "bull_case": {"type": "string"},
            "bear_case": {"type": "string"},
            "key_sources": {"type": "array", "items": {"type": "string"}}
        }
    }
)

News Monitoring

For sentiment analysis and news tracking:

# Find recent news about tracked markets
results = exa.search(
    "Bitcoin ETF SEC approval 2026",
    type="auto",
    category="news",
    start_published_date="2026-01-01T00:00:00Z",
    num_results=20,
    contents={
        "text": True,
        "highlights": {"num_sentences": 3, "highlights_per_url": 2},
    },
)

Domain Filtering for Resolution Sources

When a Kalshi market specifies a "Resolution Source", use include_domains:

# Market resolves based on BLS data
results = exa.search(
    "unemployment rate January 2026",
    include_domains=["bls.gov"],
    contents={
        "text": True,
        "livecrawl": "preferred",  # Get fresh data
    },
)

Websets API (Overview)

The Websets API enables programmatic web data discovery and processing at scale. This is a separate API with its own OpenAPI spec.

OpenAPI Spec: exa-websets-spec.yaml

Key Concepts

Concept Description
Websets Collections of web data organized around specific searches or imports
Items Individual results within a Webset
Searches Query operations that can be reused or modified
Enrichments AI-generated additional data columns added to results
Imports User-uploaded datasets for deduplication or enrichment
Monitors Scheduled operations to keep Websets updated
Webhooks Integration hooks for external systems

Use Cases

  • Data Collection: Gather web content at scale using natural language criteria
  • Data Enhancement: Add structured information via AI-powered enrichment
  • Continuous Monitoring: Keep Websets updated on scheduled intervals
  • CRM Integration: Connect results with external systems via webhooks

Note: Websets API is documented separately from the core Search/Contents APIs. See the Websets documentation for full details.

See Also