longtracer 0.1.0

RAG verification guardrails — detect hallucinations in LLM responses using hybrid STS + NLI.

LongTracer

Detect hallucinations in LLM-generated responses. LongTracer verifies every claim against source documents using hybrid STS + NLI, works with any RAG framework, and traces the full verification pipeline.

Quick Start

pip install longtracer

from longtracer import CitationVerifier

verifier = CitationVerifier()
result = verifier.verify_parallel(
    response="The Eiffel Tower is 330 meters tall and located in Berlin.",
    sources=["The Eiffel Tower is a wrought-iron lattice tower in Paris, France. It is 330 metres tall."]
)

print(result.trust_score)         # 0.0 - 1.0
print(result.hallucination_count) # 1 ("Berlin" contradicts "Paris")
print(result.all_supported)       # False

That's it. No vector store dependency, no LLM dependency. Just strings in, verification out.

How It Works

1. Claim splitting — LLM response is split into individual sentences/claims
2. STS matching — Fast bi-encoder (all-MiniLM-L6-v2) finds the best-matching source sentence for each claim
3. NLI verification — Cross-encoder (nli-deberta-v3-xsmall) classifies entailment/contradiction/neutral
4. Verdict — Trust score computed, hallucinations flagged

Framework Adapters

LangChain (3 lines)

pip install "longtracer[langchain]"

from longtracer import LongTracer, instrument_langchain

LongTracer.init(verbose=True)
instrument_langchain(your_chain)
# Your chain.invoke() now auto-verifies every response

LlamaIndex (3 lines)

pip install "longtracer[llamaindex]"

from longtracer import LongTracer, instrument_llamaindex

LongTracer.init(verbose=True)
instrument_llamaindex(your_query_engine)

Direct API (any framework)

from longtracer.guard.verifier import CitationVerifier

verifier = CitationVerifier()
result = verifier.verify_parallel(
    response="LLM said this...",
    sources=["chunk 1 text", "chunk 2 text"],
    source_metadata=[{"source": "doc.pdf", "page": 1}, {"source": "doc.pdf", "page": 2}]
)

Works with Haystack, custom pipelines, or any code that produces strings.

Multi-Project Tracing

Track multiple RAG applications independently:

from longtracer import LongTracer

LongTracer.init(project_name="chatbot-prod", backend="sqlite")

# Get project-specific tracers
chatbot = LongTracer.get_tracer("chatbot-prod")
search  = LongTracer.get_tracer("search-api")

# Each project's traces are tagged and filterable
chatbot.start_root(inputs={"query": "..."})

Vector Store & LLM Agnostic

The SDK core takes plain str and List[str]. It does not depend on any vector store (Chroma, FAISS, Pinecone, Weaviate, Qdrant, pgvector) or any LLM provider (OpenAI, Anthropic, Ollama, Bedrock). Use whatever you want — LongTracer just verifies the output.

Trace Storage Backends

LongTracer.init(backend="sqlite")   # default — persists to ~/.longtracer/traces.db
LongTracer.init(backend="memory")   # in-memory, lost on restart
LongTracer.init(backend="mongo")    # production, distributed

Backend	Install	Where traces live
SQLite	built-in (default)	~/.longtracer/traces.db
Memory	built-in	RAM only, lost on restart
MongoDB	pip install "longtracer[mongo]"	MongoDB database
PostgreSQL	pip install "longtracer[postgres]"	PostgreSQL database
Redis	pip install "longtracer[redis]"	Redis key-value store

Viewing Traces

CLI

longtracer view                        # list recent traces
longtracer view --last                 # view most recent
longtracer view --id <trace_id>        # view specific trace
longtracer view --project chatbot-prod # filter by project
longtracer view --export <trace_id>    # export to JSON
longtracer view --html <trace_id>      # export to HTML report

Console (verbose mode)

[longtracer] span=retrieval    chunks=5
[longtracer] span=llm_call     answer_len=179
[longtracer] span=eval_claims  total=3 supported=2
[longtracer] span=grounding    score=0.67 verdict=FAIL

HTML Report

from longtracer.guard.trace_report import export_trace_html
export_trace_html(tracer, filepath="report.html")

Generates a self-contained HTML file with trust score, per-claim results, timing breakdown — viewable in any browser, no external dependencies.

JSON Export

from longtracer.guard.trace_report import export_trace_json
export_trace_json(tracer, filepath="trace.json")

Optional Dependencies

Extra	Install	What it adds
langchain	pip install "longtracer[langchain]"	LangChain callback adapter
llamaindex	pip install "longtracer[llamaindex]"	LlamaIndex event adapter
mongo	pip install "longtracer[mongo]"	MongoDB trace backend
postgres	pip install "longtracer[postgres]"	PostgreSQL trace backend
redis	pip install "longtracer[redis]"	Redis trace backend
chroma	pip install "longtracer[chroma]"	ChromaDB + HuggingFace embeddings
all	pip install "longtracer[all]"	Everything

Environment Variables

Variable	Default	Description
LONGTRACER_ENABLED	false	Auto-enable with LongTracer.auto()
LONGTRACER_VERBOSE	false	Print per-span summaries
LONGTRACER_LOG_LEVEL	INFO	Python logging level
TRACE_CACHE_BACKEND	sqlite	Trace storage: sqlite, memory, mongo, postgres, redis
MONGODB_URI	—	MongoDB connection URI
POSTGRES_HOST	—	PostgreSQL host
REDIS_HOST	—	Redis host
TRACE_PROJECT	longtracer	Default project name

Demo Application

The examples/ directory contains a complete RAG demo using ChromaDB + Ollama. It is NOT part of the published PyPI package. See examples/README.md for setup instructions.

License

MIT