github/langchain
1
0
Fork 0
mirror of https://github.com/hwchase17/langchain.git synced 2025-09-13 13:36:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat(anthropic): web fetch beta (#32894)

Browse Source 
Note: citations are broken until Anthropic fixes their API
This commit is contained in:
Mason Daugherty
2025-09-11 15:14:06 -04:00
committed by GitHub
parent 83d938593b
commit 00e992a780
5 changed files with 1619 additions and 1070 deletions
Download Patch File Download Diff File Expand all files Collapse all files

96
docs/docs/integrations/chat/anthropic.ipynb
View File

@@ -970,8 +970,8 @@
"source": [
"### In tool results (agentic RAG)\n",
"\n",
":::info Requires ``langchain-anthropic>=0.3.17``\n",
"\n",
":::info\n",
"Requires ``langchain-anthropic>=0.3.17``\n",
":::\n",
"\n",
"Claude supports a [search_result](https://docs.anthropic.com/en/docs/build-with-claude/search-results) content block representing citable results from queries against a knowledge base or other custom source. These content blocks can be passed to claude both top-line (as in the above example) and within a tool result. This allows Claude to cite elements of its response using the result of a tool call.\n",
@@ -1290,6 +1290,58 @@
"print(f\"Key Points: {result.key_points}\")"
]
},
{
"cell_type": "markdown",
"id": "c580c20a",
"metadata": {},
"source": [
"### Web fetching\n",
"\n",
"Claude can use a [web fetching tool](https://docs.anthropic.com/en/docs/agents-and-tools/tool-use/web-fetch-tool) to run searches and ground its responses with citations."
]
},
{
"cell_type": "markdown",
"id": "5cf6ad08",
"metadata": {},
"source": [
":::info\n",
"Web search tool is supported since ``langchain-anthropic>=0.3.20``\n",
":::"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "c4804be1",
"metadata": {},
"outputs": [],
"source": [
"from langchain_anthropic import ChatAnthropic\n",
"\n",
"llm = ChatAnthropic(\n",
" model=\"claude-3-5-haiku-latest\",\n",
" betas=[\"web-fetch-2025-09-10\"], # Enable web fetch beta\n",
")\n",
"\n",
"tool = {\"type\": \"web_fetch_20250910\", \"name\": \"web_fetch\", \"max_uses\": 3}\n",
"llm_with_tools = llm.bind_tools([tool])\n",
"\n",
"response = llm_with_tools.invoke(\n",
" \"Please analyze the content at https://example.com/article\"\n",
")"
]
},
{
"cell_type": "markdown",
"id": "088c41d0",
"metadata": {},
"source": [
":::warning\n",
"Note: you must add the `'web-fetch-2025-09-10'` beta header to use this tool.\n",
":::"
]
},
{
"cell_type": "markdown",
"id": "1478cdc6-2e52-4870-80f9-b4ddf88f2db2",
@@ -1299,14 +1351,14 @@
"\n",
"Claude can use a [code execution tool](https://docs.anthropic.com/en/docs/agents-and-tools/tool-use/code-execution-tool) to execute Python code in a sandboxed environment.\n",
"\n",
":::info Code execution is supported since ``langchain-anthropic>=0.3.14``\n",
"\n",
":::info\n",
"Code execution is supported since ``langchain-anthropic>=0.3.14``\n",
":::"
]
},
{
"cell_type": "code",
"execution_count": 1,
"execution_count": null,
"id": "2ce13632-a2da-439f-a429-f66481501630",
"metadata": {},
"outputs": [],
@@ -1315,7 +1367,7 @@
"\n",
"llm = ChatAnthropic(\n",
" model=\"claude-sonnet-4-20250514\",\n",
" betas=[\"code-execution-2025-05-22\"],\n",
" betas=[\"code-execution-2025-05-22\"], # Enable code execution beta\n",
")\n",
"\n",
"tool = {\"type\": \"code_execution_20250522\", \"name\": \"code_execution\"}\n",
@@ -1326,6 +1378,16 @@
")"
]
},
{
"cell_type": "markdown",
"id": "a6b5e15a",
"metadata": {},
"source": [
":::warning\n",
"Note: you must add the `'code_execution_20250522'` beta header to use this tool.\n",
":::"
]
},
{
"cell_type": "markdown",
"id": "24076f91-3a3d-4e53-9618-429888197061",
@@ -1404,14 +1466,14 @@
"\n",
"Claude can use a [MCP connector tool](https://docs.anthropic.com/en/docs/agents-and-tools/mcp-connector) for model-generated calls to remote MCP servers.\n",
"\n",
":::info Remote MCP is supported since ``langchain-anthropic>=0.3.14``\n",
"\n",
":::info\n",
"Remote MCP is supported since ``langchain-anthropic>=0.3.14``\n",
":::"
]
},
{
"cell_type": "code",
"execution_count": 1,
"execution_count": null,
"id": "22fc4a89-e6d8-4615-96cb-2e117349aebf",
"metadata": {},
"outputs": [],
@@ -1423,17 +1485,17 @@
" \"type\": \"url\",\n",
" \"url\": \"https://mcp.deepwiki.com/mcp\",\n",
" \"name\": \"deepwiki\",\n",
" \"tool_configuration\": { # optional configuration\n",
" \"tool_configuration\": { # Optional configuration\n",
" \"enabled\": True,\n",
" \"allowed_tools\": [\"ask_question\"],\n",
" },\n",
" \"authorization_token\": \"PLACEHOLDER\", # optional authorization\n",
" \"authorization_token\": \"PLACEHOLDER\", # Optional authorization\n",
" }\n",
"]\n",
"\n",
"llm = ChatAnthropic(\n",
" model=\"claude-sonnet-4-20250514\",\n",
" betas=[\"mcp-client-2025-04-04\"],\n",
" betas=[\"mcp-client-2025-04-04\"], # Enable MCP beta\n",
" mcp_servers=mcp_servers,\n",
")\n",
"\n",
@@ -1443,6 +1505,16 @@
")"
]
},
{
"cell_type": "markdown",
"id": "0d6d7197",
"metadata": {},
"source": [
":::warning\n",
"Note: you must add the `'mcp-client-2025-04-04'` beta header to use this tool.\n",
":::"
]
},
{
"cell_type": "markdown",
"id": "2fd5d545-a40d-42b1-ad0c-0a79e2536c9b",

37
libs/partners/anthropic/langchain_anthropic/chat_models.py
View File

@@ -71,6 +71,10 @@ class AnthropicTool(TypedDict):
def _is_builtin_tool(tool: Any) -> bool:
"""Check if a tool is a built-in Anthropic tool.
https://docs.anthropic.com/en/docs/agents-and-tools/tool-use/overview
"""
if not isinstance(tool, dict):
return False
@@ -83,6 +87,7 @@ def _is_builtin_tool(tool: Any) -> bool:
"computer_",
"bash_",
"web_search_",
"web_fetch_",
"code_execution_",
]
return any(tool_type.startswith(prefix) for prefix in _builtin_tool_prefixes)
@@ -277,16 +282,7 @@ def _format_data_content_block(block: dict) -> dict:
def _format_messages(
messages: Sequence[BaseMessage],
) -> tuple[Union[str, list[dict], None], list[dict]]:
"""Format messages for anthropic."""
"""
[
{
"role": _message_type_lookups[m.type],
"content": [_AnthropicMessageContent(text=m.content).model_dump()],
}
for m in messages
]
"""
"""Format messages for Anthropic's API."""
system: Union[str, list[dict], None] = None
formatted_messages: list[dict] = []
merged_messages = _merge_messages(messages)
@@ -440,6 +436,7 @@ def _format_messages(
"code_execution_tool_result",
"mcp_tool_result",
"web_search_tool_result",
"web_fetch_tool_result",
):
content.append(
{
@@ -452,6 +449,7 @@ def _format_messages(
"tool_use_id",
"is_error", # for mcp_tool_result
"cache_control",
"retrieved_at", # for web_fetch_tool_result
)
},
)
@@ -1137,6 +1135,24 @@ class ChatAnthropic(BaseChatModel):
"How do I update a web app to TypeScript 5.5?"
)
.. dropdown:: Web fetch (beta)
.. code-block:: python
from langchain_anthropic import ChatAnthropic
llm = ChatAnthropic(
model="claude-3-5-haiku-latest",
betas=["web-fetch-2025-09-10"], # Enable web fetch beta
)
tool = {"type": "web_fetch_20250910", "name": "web_fetch", "max_uses": 3}
llm_with_tools = llm.bind_tools([tool])
response = llm_with_tools.invoke(
"Please analyze the content at https://example.com/article"
)
.. dropdown:: Code execution
.. code-block:: python
@@ -2232,6 +2248,7 @@ def _make_message_chunk_from_anthropic_event(
"mcp_tool_result",
"server_tool_use", # Server-side tool usage
"web_search_tool_result", # Built-in web search results
"web_fetch_tool_result", # Built-in web fetch results,
)
):
if coerce_content_to_string:

4
libs/partners/anthropic/pyproject.toml
View File

@@ -7,8 +7,8 @@ authors = []
license = { text = "MIT" }
requires-python = ">=3.9"
dependencies = [
"anthropic<1,>=0.64.0",
"langchain-core<1.0.0,>=0.3.74",
"anthropic<1,>=0.67.0",
"langchain-core<1.0.0,>=0.3.76",
"pydantic<3.0.0,>=2.7.4",
]
name = "langchain-anthropic"

257
libs/partners/anthropic/tests/integration_tests/test_chat_models.py
View File

@@ -940,6 +940,263 @@ def test_web_search() -> None:
)
# @pytest.mark.vcr
@pytest.mark.xfail(reason="Citations broken in Anthropic API; all other features work")
def test_web_fetch() -> None:
"""Note: this is a beta feature.
TODO: Update to remove beta once it's generally available.
"""
llm = ChatAnthropic(model="claude-3-5-haiku-latest", betas=["web-fetch-2025-09-10"]) # type: ignore[call-arg]
tool = {"type": "web_fetch_20250910", "name": "web_fetch", "max_uses": 1}
llm_with_tools = llm.bind_tools([tool])
input_message = {
"role": "user",
"content": [
{
"type": "text",
"text": "Fetch the content at https://docs.langchain.com and analyze",
},
],
}
response = llm_with_tools.invoke([input_message])
assert all(isinstance(block, dict) for block in response.content)
block_types = {
block["type"] for block in response.content if isinstance(block, dict)
}
# A successful fetch call should include:
# 1. text response from the model (e.g. "I'll fetch that for you")
# 2. server_tool_use block indicating the tool was called (using tool "web_fetch")
# 3. web_fetch_tool_result block with the results of said fetch
assert block_types == {"text", "server_tool_use", "web_fetch_tool_result"}
# Verify web fetch result structure
web_fetch_results = [
block
for block in response.content
if isinstance(block, dict) and block.get("type") == "web_fetch_tool_result"
]
assert len(web_fetch_results) == 1 # Since max_uses=1
fetch_result = web_fetch_results[0]
assert "content" in fetch_result
assert "url" in fetch_result["content"]
assert "retrieved_at" in fetch_result["content"]
# Fetch with citations enabled
tool_with_citations = tool.copy()
tool_with_citations["citations"] = {"enabled": True}
llm_with_citations = llm.bind_tools([tool_with_citations])
citation_message = {
"role": "user",
"content": (
"Fetch https://docs.langchain.com and provide specific quotes with "
"citations"
),
}
citation_response = llm_with_citations.invoke([citation_message])
citation_results = [
block for block in citation_response.content if isinstance(block, dict)
]
assert len(citation_results) == 1 # Since max_uses=1
citation_result = citation_results[0]
assert citation_result["content"]["content"]["citations"]["enabled"]
text_blocks = [
block
for block in citation_response.content
if isinstance(block, dict) and block.get("type") == "text"
]
# Check that the response contains actual citations in the content
has_citations = False
for block in text_blocks:
citations = block.get("citations", [])
for citation in citations:
if citation.get("type") and citation.get("start_char_index"):
has_citations = True
break
assert has_citations, (
"Expected inline citation tags in response when citations are enabled for "
"web fetch"
)
# Max content tokens param
tool_with_limit = tool.copy()
tool_with_limit["max_content_tokens"] = 1000
llm_with_limit = llm.bind_tools([tool_with_limit])
limit_response = llm_with_limit.invoke([input_message])
# Response should still work even with content limits
assert any(
block["type"] == "web_fetch_tool_result"
for block in limit_response.content
if isinstance(block, dict)
)
# Domains filtering (note: only one can be set at a time)
tool_with_allowed_domains = tool.copy()
tool_with_allowed_domains["allowed_domains"] = ["docs.langchain.com"]
llm_with_allowed = llm.bind_tools([tool_with_allowed_domains])
allowed_response = llm_with_allowed.invoke([input_message])
assert any(
block["type"] == "web_fetch_tool_result"
for block in allowed_response.content
if isinstance(block, dict)
)
# Test that a disallowed domain doesn't work
tool_with_disallowed_domains = tool.copy()
tool_with_disallowed_domains["allowed_domains"] = [
"example.com"
] # Not docs.langchain.com
llm_with_disallowed = llm.bind_tools([tool_with_disallowed_domains])
disallowed_response = llm_with_disallowed.invoke([input_message])
# We should get an error result since the domain (docs.langchain.com) is not allowed
disallowed_results = [
block
for block in disallowed_response.content
if isinstance(block, dict) and block.get("type") == "web_fetch_tool_result"
]
if disallowed_results:
disallowed_result = disallowed_results[0]
if disallowed_result.get("content", {}).get("type") == "web_fetch_tool_error":
assert disallowed_result["content"]["error_code"] in [
"invalid_url",
"fetch_failed",
]
# Blocked domains filtering
tool_with_blocked_domains = tool.copy()
tool_with_blocked_domains["blocked_domains"] = ["example.com"]
llm_with_blocked = llm.bind_tools([tool_with_blocked_domains])
blocked_response = llm_with_blocked.invoke([input_message])
assert any(
block["type"] == "web_fetch_tool_result"
for block in blocked_response.content
if isinstance(block, dict)
)
# Test fetching from a blocked domain fails
blocked_domain_message = {
"role": "user",
"content": "Fetch https://example.com and analyze",
}
tool_with_blocked_example = tool.copy()
tool_with_blocked_example["blocked_domains"] = ["example.com"]
llm_with_blocked_example = llm.bind_tools([tool_with_blocked_example])
blocked_domain_response = llm_with_blocked_example.invoke([blocked_domain_message])
# Should get an error when trying to access a blocked domain
blocked_domain_results = [
block
for block in blocked_domain_response.content
if isinstance(block, dict) and block.get("type") == "web_fetch_tool_result"
]
if blocked_domain_results:
blocked_result = blocked_domain_results[0]
if blocked_result.get("content", {}).get("type") == "web_fetch_tool_error":
assert blocked_result["content"]["error_code"] in [
"invalid_url",
"fetch_failed",
]
# Max uses parameter - test exceeding the limit
multi_fetch_message = {
"role": "user",
"content": (
"Fetch https://docs.langchain.com and then try to fetch "
"https://python.langchain.com"
),
}
max_uses_response = llm_with_tools.invoke([multi_fetch_message])
# Should contain at least one fetch result and potentially an error for the second
fetch_results = [
block
for block in max_uses_response.content
if isinstance(block, dict) and block.get("type") == "web_fetch_tool_result"
] # type: ignore[index]
assert len(fetch_results) >= 1
error_results = [
r
for r in fetch_results
if r.get("content", {}).get("type") == "web_fetch_tool_error"
]
if error_results:
assert any(
r["content"]["error_code"] == "max_uses_exceeded" for r in error_results
)
# Streaming
full: Optional[BaseMessageChunk] = None
for chunk in llm_with_tools.stream([input_message]):
assert isinstance(chunk, AIMessageChunk)
full = chunk if full is None else full + chunk
assert isinstance(full, AIMessageChunk)
assert isinstance(full.content, list)
block_types = {block["type"] for block in full.content if isinstance(block, dict)}
assert block_types == {"text", "server_tool_use", "web_fetch_tool_result"}
# Test that URLs from context can be used in follow-up
next_message = {
"role": "user",
"content": "What does the site you just fetched say about models?",
}
follow_up_response = llm_with_tools.invoke(
[input_message, full, next_message],
)
# Should work without issues since URL was already in context
assert isinstance(follow_up_response.content, (list, str))
# Error handling - test with an invalid URL format
error_message = {
"role": "user",
"content": "Try to fetch this invalid URL: not-a-valid-url",
}
error_response = llm_with_tools.invoke([error_message])
# Should handle the error gracefully
assert isinstance(error_response.content, (list, str))
# PDF document fetching
pdf_message = {
"role": "user",
"content": (
"Fetch this PDF: "
"https://www.w3.org/WAI/ER/tests/xhtml/testfiles/resources/pdf/dummy.pdf "
"and summarize its content",
),
}
pdf_response = llm_with_tools.invoke([pdf_message])
assert any(
block["type"] == "web_fetch_tool_result"
for block in pdf_response.content
if isinstance(block, dict)
)
# Verify PDF content structure (should have base64 data for PDFs)
pdf_results = [
block
for block in pdf_response.content
if isinstance(block, dict) and block.get("type") == "web_fetch_tool_result"
]
if pdf_results:
pdf_result = pdf_results[0]
content = pdf_result.get("content", {})
if content.get("content", {}).get("source", {}).get("type") == "base64":
assert content["content"]["source"]["media_type"] == "application/pdf"
assert "data" in content["content"]["source"]
@pytest.mark.vcr
def test_code_execution() -> None:
"""Note: this is a beta feature.

2295
libs/partners/anthropic/uv.lock generated
View File

File diff suppressed because it is too large Load Diff