LlamaIndex integration

Use QWEDQueryEngine wrapper for LlamaIndex to add verification to RAG pipelines. Includes math and fact checking for retrieval-augmented generation workflows.

​

Installation

pip install qwed llama-index

​

Quick start

from qwed_sdk.llamaindex import QWEDQueryEngine

# Wrap any query engine with verification
verified_engine = QWEDQueryEngine(base_engine)

response = verified_engine.query("What is 15% of 200?")
print(response.verified)  # True/False
print(response.response)  # The answer

​

QWEDQueryEngine

Wraps any LlamaIndex query engine to add verification:

from llama_index.core import VectorStoreIndex
from qwed_sdk.llamaindex import QWEDQueryEngine

# Create base engine
index = VectorStoreIndex.from_documents(documents)
base_engine = index.as_query_engine()

# Wrap with QWED
verified_engine = QWEDQueryEngine(
    base_engine,
    api_key="qwed_...",
    verify_math=True,
    verify_facts=True,
    auto_correct=False,
)

response = verified_engine.query("Calculate the total cost")
print(response.verified)

​

How responses are verified

Changed in v8.0.0 (GHSA-83v6-6h23-pvp4)

response.verified is True only when a QWED engine checked the response and the server admitted it. The integration reads the client’s is_verified, which follows the server’s admission. Everything else fails closed:

Case	verified	status
A math response the server admits	True	Server status, for example VERIFIED
A math response the server does not admit	False	Server status, for example BLOCKED
No QWED engine applies, or a fact query without source context	False	UNVERIFIABLE
The check raises an error	False	ERROR

response = verified_engine.query("Where is the Eiffel Tower?")

if not response.verified:
    # UNVERIFIABLE, ERROR, or not admitted by the server
    print(response.status)

Before v8.0.0, QWEDQueryEngine returned verified=True with status PASSED for any response no engine checked, including fact-style queries. Don’t rely on verified=True from earlier versions.

​

QWEDVerificationTransform

Node postprocessor that verifies retrieved content:

from qwed_sdk.llamaindex import QWEDVerificationTransform

engine = index.as_query_engine(
    node_postprocessors=[
        QWEDVerificationTransform(
            verify_math=True,
            verify_code=True,
        )
    ]
)

The transform sets qwed_verified in each node’s metadata:

• True only when a check ran and the server admitted the content.
• False when the server refused the content or the check raised an error.
• False together with qwed_checked: False when the node had nothing to check.

​

QWEDCallbackHandler

Track verification across all operations:

from llama_index.core import Settings
from qwed_sdk.llamaindex import QWEDCallbackHandler

Settings.callback_manager.add_handler(
    QWEDCallbackHandler(log_all=True)
)

​

QWEDVerifyTool

For LlamaIndex agents:

from llama_index.core.agent import ReActAgent
from qwed_sdk.llamaindex import QWEDVerifyTool

tools = [QWEDVerifyTool()]
agent = ReActAgent.from_tools(tools, llm=llm)

​

VerifiedResponse

@dataclass
class VerifiedResponse:
    response: str
    verified: bool
    status: str
    confidence: float
    attestation: Optional[str]
    source_nodes: List[Any]