Contents API

Extract clean, LLM-ready web content.

Just want working code?

Stop reading. Visit contents coding agent reference and copy paste to your agent.

What it is

/contents returns clean, structured content from any URL, handling JavaScript-rendered pages, PDFs, and complex layouts automatically. You pass in URLs and choose full page text, targeted highlights, or LLM-generated summaries. It can also crawl linked subpages to pull content from entire site sections in a single request.

All contents features are also available in /search for returned URLs, at no extra charge up to 10 results per search ($1/1000 pages afterwards). We recommend using /search in this way instead of /contents for web search tool use cases.

Use /contents when you already know the URLs. If you are starting from a query and want Exa to find the pages first, start with Search.

Key capabilities

Content modes

Choose the content view that matches the task:

Mode	What You Get	Best For
Text	Full page content as clean markdown	Deep analysis, full context research
Highlights	Key excerpts relevant to your query	Per-page evidence and factual lookups
Dynamic Highlights	Excerpts allocated across the result set	Shared agent or RAG context
Summary	LLM-generated abstract	Quick overviews, structured extraction with JSON schema

Subpage crawling

Automatically discover and extract content from linked pages within a site. Pass subpages: 10 and optionally subpageTarget: ["docs", "about"] to focus on relevant sections.

Content freshness

Control whether results come from cache or are freshly crawled with maxAgeHours:

Setting	Behavior
Omit (default)	Livecrawl only when no cache exists
24	Use cache if < 24 hours old, otherwise livecrawl
0	Always livecrawl (slowest, freshest)
-1	Cache only (fastest, may be stale)

Dynamic highlights

Dynamic Highlights is available as a research preview on /search and /contents. Include the Exa-Beta: dynamic-highlights-2026-08-28 header on every request that sets dynamic: true.

Regular highlights find relevant excerpts within each page independently. Dynamic Highlights considers the pages together and allocates one shared context budget across the result set. Useful pages can receive more context, while redundant or weak pages can receive less context.

Use it when several pages will feed the same agent or RAG context. Keep regular highlights when every page needs its own excerpt or a predictable per-page limit.

Search

On /search, enable it inside contents.highlights:

curl -X POST 'https://api.exa.ai/search' \
  -H "x-api-key: $EXA_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Exa-Beta: dynamic-highlights-2026-08-28' \
  -d '{
    "query": "How are inference providers reducing transformer latency?",
    "numResults": 5,
    "contents": {
      "highlights": {
        "dynamic": true
      }
    }
  }'

Response:

{
  "results": [
    {
      "title": "How to optimize LLM inference speed and reduce costs in production",
      "url": "https://www.baseten.co/blog/how-to-optimize-llm-inference-speed-and-reduce-costs-in-production/",
      "highlights": [
        "During one decode iteration, the GPU generates one token for every active request in the batch. The problem with traditional batching is that the server waits for every request in the batch to finish before accepting new ones. ... Speculative decoding lets you generate multiple tokens per decode step ..."
      ]
    },
    {
      "title": "Smaller, faster, safer: running Kimi and GLM at scale",
      "url": "https://blog.cloudflare.com/smaller-faster-safer-models/",
      "highlights": [
        "... separating the prefill and decode phases of inference to get more out of each GPU. This post looks at three techniques we layer on top of that to fit these models into memory and keep them fast: quantizing the KV cache, compressing the model weights ..."
      ]
    },
    {
      "title": "How Modern LLM Inference Became 10-100x Faster",
      "url": "https://nandigamharikrishna.substack.com/p/how-modern-llm-inference-became-10100x",
      "highlights": [
        "... In the vLLM paper, PagedAttention achieved near-zero KV-cache ... 2 to 4x throughput improvements over systems such as FasterTransformer and Orca at similar latency."
      ]
    }
  ]
}

The most useful pages above received several thousand characters of the shared budget, while thinner pages received a few hundred.

Contents

On /contents, highlights remains a top-level field:

curl -X POST 'https://api.exa.ai/contents' \
  -H "x-api-key: $EXA_API_KEY" \
  -H 'Content-Type: application/json' \
  -H 'Exa-Beta: dynamic-highlights-2026-08-28' \
  -d '{
    "urls": [
      "https://www.baseten.co/blog/how-to-optimize-llm-inference-speed-and-reduce-costs-in-production/",
      "https://blog.cloudflare.com/smaller-faster-safer-models/",
      "https://www.crusoe.ai/resources/blog/430-tokens-per-second-optimizing-kimi-k2-6-and-k2-7-for-production"
    ],
    "highlights": {
      "dynamic": true,
      "query": "How are inference providers reducing transformer latency?"
    }
  }'

Response:

{
  "results": [
    {
      "url": "https://www.baseten.co/blog/how-to-optimize-llm-inference-speed-and-reduce-costs-in-production/",
      "highlights": [
        "A batch is a group of requests processed together on the GPU at the same time. Batching matters because GPUs are built to handle multiple computations from different requests in parallel. ..."
      ]
    },
    {
      "url": "https://blog.cloudflare.com/smaller-faster-safer-models/",
      "highlights": [
        "We've written before about how we serve large models on Workers AI and about separating the prefill and decode phases of inference to get more out of each GPU. ..."
      ]
    },
    {
      "url": "https://www.crusoe.ai/resources/blog/430-tokens-per-second-optimizing-kimi-k2-6-and-k2-7-for-production",
      "highlights": [
        "Through rigorous profiling, we identified a decode kernel that was operating suboptimally for specific Kimi workload shapes. We developed a custom optimization for this path ... This change alone added approximately 40 output tokens per second. ..."
      ]
    }
  ]
}

The response shape does not change: each result still has a highlights array.

Do not combine dynamic: true with maxCharacters. Dynamic Highlights sizes and distributes the shared output budget automatically.

Common use cases

Token-efficient info from an article

Get the most relevant excerpts without needing the full page.

result = exa.get_contents(
  ["https://example.com/research-paper"],
  highlights={"query": "methodology and results"}
)

Structured outputs using summaries

Extract specific fields from any page using a JSON schema.

result = exa.get_contents(
  ["https://example.com/company-page"],
  summary={
    "query": "Extract company information",
    "schema": {
      "type": "object",
      "properties": {
        "name": {"type": "string"},
        "industry": {"type": "string"},
        "founded": {"type": "number"}
      },
      "required": ["name", "industry"]
    }
  }
)

Crawl many pages from a website

Pull content from a docs site, targeting specific sections.

result = exa.get_contents(
  ["https://docs.example.com"],
  subpages=15,
  subpage_target=["api", "models", "embeddings"],
  max_age_hours=24,
  text={"max_characters": 5000}
)

Human Quickstart

Get your API key from the Exa Dashboard, then set it as an environment variable:

macOS/Linux

export EXA_API_KEY="your-api-key"

Windows

setx EXA_API_KEY "your-api-key"

Install the SDK:

Python

pip install exa-py

Then make your first request:

Python

from exa_py import Exa

exa = Exa()

result = exa.get_contents(
  ["https://example.com/article"],
  highlights=True
)

Next

• Search API - Find content on the web with natural language
• Contents API Reference - Full API reference with all parameters
• MCP Setup - Connect your AI assistant to Exa
• SDKs - Python and JavaScript SDK docs