langchain-truthvouch

LangChain integration for the TruthVouch Trust API. Add hallucination detection and content verification to any LangChain pipeline in minutes.

Installation

pip install langchain-truthvouch

Quickstart

import os
os.environ["TRUTHVOUCH_API_KEY"] = "your-api-key"

from truthvouch_langchain import TruthVouchGuard

guard = TruthVouchGuard(threshold=0.8)
result = guard("The Eiffel Tower is located in Berlin.")
print(result.passed)        # False — low trust score
print(result.trust_score)   # e.g. 0.21
print(result.claims)        # list of per-claim verification results

Components

TrustApiClient

Low-level async/sync client for the Trust API verify endpoint.

from truthvouch_langchain import TrustApiClient

client = TrustApiClient(
    api_key="your-api-key",           # or set TRUTHVOUCH_API_KEY env var
    base_url="http://localhost:5004/api/v1/trust",  # default
)

result = client.verify_sync("Paris is the capital of France.", mode="standard")
print(result.trust_score)  # 0.97

TruthVouchRetriever

Wraps any document list with trust verification. Documents below the threshold are filtered out.

from langchain_core.documents import Document
from truthvouch_langchain import TruthVouchRetriever

retriever = TruthVouchRetriever(
    api_key="your-api-key",
    threshold=0.75,
    mode="spot_check",
)
docs = retriever.get_relevant_documents("What is the capital of France?")
# Each doc has doc.metadata["trust_score"] set

TruthVouchCallbackHandler

Automatically verifies every LLM response via a LangChain callback.

from langchain_openai import ChatOpenAI
from truthvouch_langchain import TruthVouchCallbackHandler

handler = TruthVouchCallbackHandler(
    api_key="your-api-key",
    threshold=0.8,
    raise_on_low_trust=False,  # set True to raise TrustThresholdError
)

llm = ChatOpenAI(callbacks=[handler])
response = llm.invoke("Tell me about the Eiffel Tower.")

TruthVouchGuard

Simple callable guard — verify any text before displaying or acting on it.

from truthvouch_langchain import TruthVouchGuard

guard = TruthVouchGuard(threshold=0.8)
result = guard("Some text to verify")
if not result.passed:
    print("Low trust score:", result.trust_score)

Configuration

License

Apache-2.0