LangChain docs:https://docs.langchain.com/oss/python/integrations/providers/diffbot
A thin LangChain integration over the official diffbot-python SDK. Every Diffbot API gets the closest LangChain primitive:
| Diffbot API | LangChain class(es) |
|---|---|
| Knowledge Graph (DQL) | DiffbotKnowledgeGraphRetriever, DiffbotKnowledgeGraphTool |
| Web Search | DiffbotWebSearchRetriever, DiffbotWebSearchTool |
| Extract (Analyze) | DiffbotExtractTool, DiffbotExtractLoader |
| NLP entities | DiffbotEntitiesTool |
| Crawl | DiffbotCrawlLoader |
LLM RAG (ask) | ChatDiffbot (with native streaming), DiffbotAskTool |
pip install langchain-diffbotGet an API token at https://app.diffbot.com/get-started/.
Every component takes a pre-built SDK client — you build a diffbot.Diffbot (sync) and/or diffbot.DiffbotAsync (async) and pass it via client= / async_client=. That's the only way to give a component HTTP access, and it keeps configuration in one place: customize the client (token, timeout, transport=, custom URLs) however the SDK allows, and share one client across many components to reuse a single connection pool. The component uses the client as-is and never closes it — you own its lifecycle.
importosfromdiffbotimportDiffbotdb=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"])Pick your execution mode by which client you build: Diffbot for the sync surface (invoke, stream, load), DiffbotAsync for the async surface (ainvoke, astream, alazy_load). Pass both if a component is used both ways.
importosfromdiffbotimportDiffbotfromlangchain_diffbotimportDiffbotKnowledgeGraphRetrieverdb=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"])
retriever=DiffbotKnowledgeGraphRetriever(client=db, k=5)
docs=retriever.invoke("type:Organization industries:\"Artificial Intelligence\" location.city.name:\"Boston\"")
fordindocs:
print(d.metadata["name"], "—", d.page_content[:120])The query string is a DQL (Diffbot Query Language) expression.
Diffbot KG entities and web-search results are large. Dumping them straight into an LLM prompt can blow past per-minute input-token limits in a single call. Both retrievers expose three shaping knobs:
importosfromdiffbotimportDiffbotfromlangchain_core.documentsimportDocumentfromlangchain_diffbotimportDiffbotKnowledgeGraphRetrieverdb=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"])
# 1. Project only the top-level fields you care about. Drops everything else# from `metadata`. Recommended for agent / tool-use scenarios.retriever=DiffbotKnowledgeGraphRetriever(
client=db,
k=5,
fields=["id", "type", "name", "homepageUri", "nbEmployees"],
)
# 2. Choose which field becomes `page_content`. First non-empty value wins.retriever=DiffbotKnowledgeGraphRetriever(
client=db,
content_fields=["summary", "description", "name"],
)
# 3. For total control, pass a `document_mapper` that turns a raw entity# dict into whatever Document shape you want.defmapper(entity: dict) ->Document:
returnDocument(
page_content=entity.get("summary", ""),
metadata={"id": entity["id"], "name": entity["name"]},
)
retriever=DiffbotKnowledgeGraphRetriever(client=db, document_mapper=mapper)fields and content_fields are ignored when document_mapper is set. The same knobs work on DiffbotWebSearchRetriever.
importosfromdiffbotimportDiffbotfromlangchain_diffbotimportDiffbotWebSearchRetrieverdb=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"])
web=DiffbotWebSearchRetriever(client=db, k=5, fields=["title", "pageUrl", "score"])
docs=web.invoke("diffbot knowledge graph llm grounding")importosfromdiffbotimportDiffbotfromlangchain_diffbotimportDiffbotExtractTool, DiffbotExtractLoaderdb=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"])
# Single URLtool=DiffbotExtractTool(client=db)
page=tool.invoke({"url": "https://www.diffbot.com/products/extract/"})
# Batch — yields one Document per URL (the same client is reused)loader=DiffbotExtractLoader(client=db, urls=["https://example.com", "https://diffbot.com"])
fordocinloader.lazy_load():
print(doc.metadata["title"], doc.page_content[:200])DiffbotExtractTool returns a structured {"error": ..., "errorCode": ...} dict when Diffbot reports an extraction failure (200 with errorCode), so agents can react and try another URL instead of catching an exception. Auth / rate-limit errors propagate as diffbot.errors.AuthError / RateLimitError.
DiffbotCrawlLoader drives a Diffbot crawl job and yields one Document per crawled URL. The page_content is the URL itself (the crawl API surfaces URLs, not page contents) — chain it with DiffbotExtractLoader to fetch the content of each URL.
importosfromdiffbotimportDiffbotfromlangchain_diffbotimportDiffbotCrawlLoaderdb=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"])
loader=DiffbotCrawlLoader(client=db, site="https://www.diffbot.com")
fordocinloader.lazy_load():
print(doc.metadata["url"], doc.metadata["status"])importosfromdiffbotimportDiffbotfromlangchain.messagesimportHumanMessagefromlangchain_diffbotimportChatDiffbotllm=ChatDiffbot(client=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"]))
forchunkinllm.stream([HumanMessage(content="What is the Diffbot Knowledge Graph?")]):
print(chunk.content, end="", flush=True)_stream / _astream are native — no thread-pool fallback. .invoke() aggregates the stream into a single message.
To let a tool-calling agent consult Diffbot's LLM (rather than use it as the primary model), hand it DiffbotAskTool instead — it answers a natural-language question from the Knowledge Graph + live web and returns a synthesized string:
importosfromdiffbotimportDiffbotfromlangchain_diffbotimportDiffbotAskToolask=DiffbotAskTool(client=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"]))
print(ask.invoke({"question": "Who founded Diffbot, and when?"}))Every Diffbot API is also exposed as an agent-callable BaseTool. Hand a tool-calling agent only the tools you want — they all share whatever client you pass. DiffbotExtractTool and DiffbotAskTool are shown above; the rest:
Runs a Diffbot web search and returns the result list — each item with title, pageUrl, score, and content. New accounts include 100,000 free web searches per month. (Use DiffbotWebSearchRetriever when you want Document output instead.)
importosfromdiffbotimportDiffbotfromlangchain_diffbotimportDiffbotWebSearchTooltool=DiffbotWebSearchTool(client=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"]))
results=tool.invoke({"text": "diffbot knowledge graph"})Runs a DQL query against the Knowledge Graph from within an agent and returns the raw response dict. (Use DiffbotKnowledgeGraphRetriever when you want Document output instead.)
importosfromdiffbotimportDiffbotfromlangchain_diffbotimportDiffbotKnowledgeGraphTooltool=DiffbotKnowledgeGraphTool(client=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"]))
body=tool.invoke({"query": 'type:Organization name:"Diffbot"'})Identifies named entities and sentiment in text via Diffbot's NLP API. The returned entity IDs can be looked up in the Knowledge Graph (e.g. id:or("E1","E2")).
importosfromdiffbotimportDiffbotfromlangchain_diffbotimportDiffbotEntitiesTooltool=DiffbotEntitiesTool(client=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"]))
result=tool.invoke({"text": "Diffbot was founded in Menlo Park."})So an agent can build valid DQL instead of guessing field names, two tools wrap Diffbot's DQL-authoring helpers. The intended loop is introspect (ontology) → probe → run (DiffbotKnowledgeGraphTool) → refine.
DiffbotOntologyTool navigates the KG ontology — discover real entity types, field paths, taxonomy, and enum values before querying. The ontology is fetched once over HTTP and cached on the tool instance for the rest of its lifetime.
importosfromdiffbotimportDiffbotfromlangchain_diffbotimportDiffbotOntologyTooltool=DiffbotOntologyTool(client=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"]))
types=tool.invoke({"op": "types"})DiffbotDQLProbeTool probes query variants at size=0 (hit counts only), so an agent can check selectivity — not zero, not millions — before committing to a full query.
importosfromdiffbotimportDiffbotfromlangchain_diffbotimportDiffbotDQLProbeTooltool=DiffbotDQLProbeTool(client=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"]))
counts=tool.invoke({"queries": ['type:Organization name:"Diffbot"', "type:Person"]})The retrievers are standard BaseRetrievers, so they slot into LCEL like any other:
importosfromdiffbotimportDiffbotfromlangchain_anthropicimportChatAnthropicfromlangchain_core.output_parsersimportStrOutputParserfromlangchain_core.promptsimportChatPromptTemplatefromlangchain_core.runnablesimportRunnablePassthroughfromlangchain_diffbotimportDiffbotKnowledgeGraphRetrieverretriever=DiffbotKnowledgeGraphRetriever(
client=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"]),
k=5,
fields=["id", "name", "homepageUri", "nbEmployees", "industries"],
)
prompt=ChatPromptTemplate.from_template(
"Answer using only this Diffbot KG context:\n\n{context}\n\nQuestion: {question}"
)
def_format(docs):
return"\n---\n".join(
f"{d.metadata.get('name')} (id={d.metadata.get('id')}): {d.page_content}"fordindocs
)
chain= (
{"context": retriever|_format, "question": RunnablePassthrough()}
|prompt|ChatAnthropic(model="claude-sonnet-4-6")
|StrOutputParser()
)
chain.invoke('type:Organization location.city.name:"Boston" industries:"Biotech"')Because every component takes a client, you configure the SDK once and hand the same client to as many components as you like — they share its connection pool, and there's no per-call pool churn. Build the tools/retrievers you actually want and add only those to your agent; the client is the shared resource, not a bundle.
importosfromdiffbotimportDiffbotfromlangchain_diffbotimport (
DiffbotKnowledgeGraphTool,
DiffbotAskTool,
DiffbotWebSearchRetriever,
)
# One client, configured once (timeout, transport, custom URLs, ...),# shared across every component.db=Diffbot(token=os.environ["DIFFBOT_API_TOKEN"], timeout=60.0)
kg=DiffbotKnowledgeGraphTool(client=db)
ask=DiffbotAskTool(client=db)
web=DiffbotWebSearchRetriever(client=db, k=5)
# `db.close()` when you're done — the components never close it for you.Anything the SDK supports (custom URLs, transport=, headers via a custom transport) is configured on the client you build — there's no second configuration surface to learn. For async, build a diffbot.DiffbotAsync and pass async_client= instead (or both, if a component is used both ways).
| Class | Abstraction | Import path |
|---|---|---|
ChatDiffbot | Chat model | from langchain_diffbot import ChatDiffbot |
DiffbotKnowledgeGraphRetriever | Retriever | from langchain_diffbot import DiffbotKnowledgeGraphRetriever |
DiffbotWebSearchRetriever | Retriever | from langchain_diffbot import DiffbotWebSearchRetriever |
DiffbotExtractLoader | Document loader | from langchain_diffbot import DiffbotExtractLoader |
DiffbotCrawlLoader | Document loader | from langchain_diffbot import DiffbotCrawlLoader |
DiffbotExtractTool | Tool | from langchain_diffbot import DiffbotExtractTool |
DiffbotWebSearchTool | Tool | from langchain_diffbot import DiffbotWebSearchTool |
DiffbotKnowledgeGraphTool | Tool | from langchain_diffbot import DiffbotKnowledgeGraphTool |
DiffbotEntitiesTool | Tool | from langchain_diffbot import DiffbotEntitiesTool |
DiffbotAskTool | Tool | from langchain_diffbot import DiffbotAskTool |
DiffbotOntologyTool | Tool | from langchain_diffbot import DiffbotOntologyTool |
DiffbotDQLProbeTool | Tool | from langchain_diffbot import DiffbotDQLProbeTool |
The examples/ folder has runnable demos:
examples/quickstart/— full tour: every public class, output shaping, async, and a multi-tool research agent.examples/company_research/— the same multi-tool agent as a one-shot CLI:cd examples && python -m company_research "your question". The agent combines KG search + web search + URL extract.
Both need langchain + langchain-anthropic on top of the base package — install the extra:
pip install "langchain-diffbot[examples]"uv sync --all-groups
uv run pytest tests/unit_testsIntegration tests hit the live Diffbot API and require DIFFBOT_API_TOKEN:
uv run pytest tests/integration_tests