Documentation Index
Fetch the complete documentation index at: https://docs.langchain.com/llms.txt
Use this file to discover all available pages before exploring further.
Parallel is a real-time web search and content extraction platform built for LLMs and AI applications.
ParallelExtractTool calls Parallel’s Extract API, which returns clean, markdown-formatted content from web pages, with optional focused excerpts driven by a search_objective. Pair it with ParallelSearchTool to build a search → extract pipeline.
Overview
Integration details
| Class | Package | Serializable | JS support | Package latest |
|---|
ParallelExtractTool | langchain-parallel | ❌ | ❌ |  |
Setup
The integration lives in the langchain-parallel package.
pip install -U langchain-parallel
Credentials
Head to Parallel to sign up and generate an API key. Set PARALLEL_API_KEY in your environment:
import getpass
import os
if not os.environ.get("PARALLEL_API_KEY"):
os.environ["PARALLEL_API_KEY"] = getpass.getpass("Parallel API key:\n")
Instantiation
from langchain_parallel import ParallelExtractTool
tool = ParallelExtractTool()
# Or pass an explicit key, override the base URL, or cap per-URL `full_content` size:
# tool = ParallelExtractTool(
# api_key="your-api-key",
# base_url="https://api.parallel.ai",
# max_chars_per_extract=5000,
# )
Invocation
Invoke directly with args
result = tool.invoke(
{"urls": ["https://en.wikipedia.org/wiki/Artificial_intelligence"]}
)
print(result[0]["title"])
print(result[0]["url"])
print(result[0]["content"][:200], "...")
result = tool.invoke(
{
"urls": [
"https://en.wikipedia.org/wiki/Machine_learning",
"https://en.wikipedia.org/wiki/Deep_learning",
"https://en.wikipedia.org/wiki/Natural_language_processing",
]
}
)
for item in result:
print(item["title"], "—", item["url"])
ToolCall returns a ToolMessage:
model_generated_tool_call = {
"args": {
"urls": [
"https://en.wikipedia.org/wiki/Climate_change",
"https://en.wikipedia.org/wiki/Renewable_energy",
]
},
"id": "call_123",
"name": tool.name, # "parallel_extract"
"type": "tool_call",
}
result = tool.invoke(model_generated_tool_call)
Async usage
async def extract_async():
return await tool.ainvoke(
{
"urls": [
"https://en.wikipedia.org/wiki/Python_(programming_language)",
"https://en.wikipedia.org/wiki/JavaScript",
]
}
)
result = await extract_async()
Focused excerpts
Drive excerpt selection with a search_objective (or search_queries). Setting full_content=False skips the full markdown body and returns only matched excerpts:
result = tool.invoke(
{
"urls": ["https://en.wikipedia.org/wiki/Artificial_intelligence"],
"search_objective": "What are the main applications and ethical concerns of AI?",
"excerpts": {"max_chars_per_result": 2000},
"full_content": False,
}
)
Fetch policy and full-content sizing
Control caching, timeouts, and the per-URL full_content cap independently:
result = tool.invoke(
{
"urls": ["https://en.wikipedia.org/wiki/Quantum_computing"],
"fetch_policy": {
"max_age_seconds": 86400,
"timeout_seconds": 60,
"disable_cache_fallback": False,
},
"full_content": {"max_chars_per_result": 5000},
}
)
full_content precedence. An explicit FullContentSettings (or dict) on the call always wins over the tool-level max_chars_per_extract. The latter only applies when you pass full_content=True as a plain bool.
Per-URL error handling
Failed URLs are returned as items with error_type set, so partial-success is the default:
result = tool.invoke(
{
"urls": [
"https://en.wikipedia.org/wiki/Artificial_intelligence",
"https://this-domain-does-not-exist-12345.com/",
]
}
)
for item in result:
if "error_type" in item:
print("failed:", item["url"], "—", item["content"])
else:
print("ok:", item["url"], f"({len(item['content'])} chars)")
Parameters
Required
urls: list of URLs to extract.
Optional
search_objective: natural-language description that drives excerpt selection.search_queries: list of keyword strings used together with (or in place of) search_objective.excerpts: per-result excerpt settings. Pass ExcerptSettings(max_chars_per_result=…) (or a dict) to control per-result excerpt size; omit for the API default.full_content: True to return full markdown content (sized by the tool-level max_chars_per_extract), False to skip it, or FullContentSettings(max_chars_per_result=…) for fine-grained control.fetch_policy: cache control, e.g. {"max_age_seconds": 86400, "timeout_seconds": 60}.max_chars_total: cap on combined output length across all URLs.client_model / session_id: forwarded to Parallel for downstream attribution.
Chaining
Bind the tool to any tool-calling chat model and drive an agent with create_agent:
from langchain.agents import create_agent
from langchain.chat_models import init_chat_model
llm = init_chat_model(model="claude-opus-4-7")
agent = create_agent(model=llm, tools=[tool])
agent.invoke({"messages": [("human", "Summarize https://en.wikipedia.org/wiki/Quantum_computing")]})
ParallelSearchTool and ParallelExtractTool to the same agent. The model uses search to find URLs and extract to drill into the ones it picks.
from langchain_parallel import ParallelSearchTool
search = ParallelSearchTool()
extract = ParallelExtractTool()
agent = create_agent(model=llm, tools=[search, extract])
agent.invoke({
"messages": [
("human", "Find a recent peer-reviewed paper on net-energy-gain fusion and summarize it."),
]
})
[
{
"url": "https://example.com/article",
"title": "Article Title",
"content": "# Article Title\n\nMain content formatted as markdown...",
"publish_date": "2026-01-15",
"excerpts": ["...", "..."], # if excerpts/search_objective requested
},
# Failed extractions:
{
"url": "https://failed-site.com",
"title": None,
"content": "Error: 404 Not Found",
"error_type": "http_error",
},
]
API reference
For detailed documentation, head to the ParallelExtractTool API reference or the Parallel Extract reference.
