Haystack Integration

LongTracer provides a native Haystack v2 @component for verifying LLM responses in any Haystack pipeline.

Install

pip install "longtracer[haystack]"

Usage as a Component

from haystack import Pipeline, Document
from longtracer.adapters.haystack_handler import LongTracerVerifier

# Create the verifier component
verifier = LongTracerVerifier(threshold=0.5)

# Add to your pipeline
pipeline = Pipeline()
pipeline.add_component("retriever", your_retriever)
pipeline.add_component("generator", your_generator)
pipeline.add_component("verifier", verifier)

# Connect: generator output → verifier
pipeline.connect("retriever.documents", "verifier.documents")
pipeline.connect("generator.replies", "verifier.response")

Quick Instrument

from longtracer import instrument_haystack

pipeline = Pipeline()
# ... add your components ...
instrument_haystack(pipeline)

This adds a longtracer_verifier component to the pipeline. You still need to connect it to your generator and retriever outputs.

Component Inputs / Outputs

Input	Type	Description
response	str	LLM-generated text to verify
documents	List[Document]	Retrieved Haystack Documents

Output	Type	Description
response	str	Pass-through of original response
trust_score	float	0.0–1.0
verdict	str	"PASS" or "FAIL"
summary	str	Human-readable summary
claims	list	Per-claim verification results
hallucination_count	int	Number of hallucinated claims

With Tracing

from longtracer import LongTracer

LongTracer.init(verbose=True, backend="sqlite")

# The verifier component automatically picks up the active tracer
verifier = LongTracerVerifier()

Notes

• Models are loaded on first run() call (or explicitly via warm_up())
• A failing verification never crashes your pipeline — returns verdict="ERROR" with the error message
• Works with any Haystack v2 pipeline that produces documents and text responses