contexttrace 0.8.0

Local-first SDK and CLI for RAG and agent reliability tracing, citation checks, and failure diagnosis.

ContextTrace

Local-first evidence-chain debugging for RAG and AI agents.

ContextTrace shows where an answer stopped being grounded:

query -> retrieved context -> answer claims -> citations -> verdicts -> root cause

It is a Python SDK and CLI, not a hosted dashboard. Traces, reports, judge cache, and SQLite state stay local by default.

Install

pip install contexttrace
contexttrace init

Quickstart

contexttrace verify-demo unsupported_claim --report
contexttrace demo --dataset refund_policy
contexttrace report --last --open

Default local storage:

.contexttrace/contexttrace.db

Verify A RAG Trace

Create a portable trace with a query, answer, retrieved contexts, and optional citations:

{
  "query": "How long does refund processing take?",
  "answer": "Refunds are processed within 5 business days.",
  "contexts": [
    {
      "id": "policy",
      "text": "Customers may request refunds within 30 days of purchase."
    }
  ]
}

Run local evidence checks:

contexttrace inspect trace.json
contexttrace verify trace.json --report
contexttrace qa trace.json --corpus docs/ --report

ContextTrace classifies each claim as supported, partially_supported, unsupported, unverifiable, or contradicted, then explains the likely fix.

Local Verification Modes

Mode	Use When
lexical	Fast default checks with no optional dependencies.
semantic	Local paraphrase and role-aware contradiction checks.
local_ml	Offline hash-embedding similarity, optionally backed by a local SentenceTransformers model.
judge	Higher-accuracy local LLM judging through Ollama, LM Studio, vLLM, or a local OpenAI-compatible server.

Run the stronger local non-LLM verifier:

contexttrace verify trace.json --mode local_ml --report
contexttrace verify-benchmark --mode local_ml --case-set all

Optional neural local-ML support never downloads models automatically:

pip install "contexttrace[local-ml]"
set CONTEXTTRACE_LOCAL_ML_MODEL_PATH=C:\models\bge-small-en-v1.5

Run a local judge with Ollama:

set CONTEXTTRACE_JUDGE_PROVIDER=ollama
set CONTEXTTRACE_JUDGE_MODEL=llama3.1

contexttrace verify trace.json --mode judge --report
contexttrace judge-calibrate --case-set all --report

Remote judges are blocked while local_only: true is active. To use a remote judge, explicitly disable local-only mode and configure the provider/API key.

Diagnose And Regression-Test

# Find whether support existed elsewhere in the corpus.
contexttrace audit trace.json --corpus docs/ --report

# Compare a baseline and current answer after a prompt, model, or retriever change.
contexttrace compare baseline.json current.json --report

# Turn saved failures into replayable endpoint tests.
contexttrace suite create traces/failure.json --out contexttrace-suite.json
contexttrace suite run contexttrace-suite.json --endpoint http://localhost:8000/query --report

Common root causes include retrieval_miss, reranking_failure, chunking_issue, corpus_gap, answer_overreach, stale_source, citation_mismatch, and should_have_abstained.

Capture Existing Systems

Capture one live endpoint response:

contexttrace capture endpoint \
  --endpoint http://localhost:8000/query \
  --query "What is the refund policy?" \
  --answer-path $.answer \
  --contexts-path $.contexts \
  --citations-path $.citations \
  --out traces/refund_trace.json \
  --verify \
  --report

Or capture artifacts from Python:

from contexttrace import capture_rag_trace, write_rag_trace

trace = capture_rag_trace(
    query=question,
    answer=answer,
    contexts=retrieved_docs,
    metadata={"system": "support-rag"},
)
write_rag_trace(trace, "trace.json")

SDK Example

from contexttrace import ContextTrace

ct = ContextTrace(project="support-rag")

with ct.trace(query="What is the refund policy?") as trace:
    chunks = retriever.search("What is the refund policy?")
    trace.log_retrieval(chunks)
    trace.log_context(chunks[:5])

    answer = llm.generate("What is the refund policy?", chunks[:5])
    trace.log_answer(answer, usage={"total_tokens": 1200})
    trace.log_citations([
        {"claim": "Refunds are available within 30 days.", "source_chunk_id": "chunk_12"}
    ])

    result = trace.evaluate()
    print(result["failure"]["failure_type"])

Integrations

pip install "contexttrace[langchain]"
pip install "contexttrace[llamaindex]"
pip install "contexttrace[fastapi]"
pip install "contexttrace[langgraph]"
pip install "contexttrace[otel]"
pip install "contexttrace[all]"

Includes LangChain, LlamaIndex, FastAPI, LangGraph, and OpenTelemetry hooks.

Privacy

ContextTrace makes no network calls unless you point it at an endpoint or configure a judge provider. Local controls include:

• local_only: true
• log_chunk_text: false
• log_answer_text: false
• storage_path
• judge_cache_enabled: true
• judge_cache_path: .contexttrace/judge_cache.json

Limits

ContextTrace is a diagnostic tool, not a correctness proof. Claim extraction is rule-based, contradiction detection is conservative, and high-stakes outputs still need human review.

Links

• Repository: https://github.com/samarth1412/Context-Trace
• Documentation: https://github.com/samarth1412/Context-Trace/tree/main/docs
• Issues: https://github.com/samarth1412/Context-Trace/issues
• Changelog: https://github.com/samarth1412/Context-Trace/blob/main/CHANGELOG.md