contexttrace 0.6.0

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

ContextTrace

Debug RAG failures before users find them.

ContextTrace is a local-first Python SDK and CLI for evaluating existing RAG and AI agent systems. It records retrieved chunks, selected context, answer claims, citations, token usage, latency, and agent events, then writes local traces and HTML reports without requiring a hosted dashboard.

Install

pip install contexttrace
contexttrace --version
contexttrace init

Optional 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]"

Quickstart

contexttrace init
contexttrace demo --dataset refund_policy
contexttrace report --last
contexttrace doctor

By default, traces are stored locally in:

.contexttrace/contexttrace.db

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"])

BYO RAG Endpoint

Capture and verify one live response from a running local or hosted RAG API without adding SDK code:

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

If you already have a saved endpoint response:

contexttrace capture response response.json \
  --query "What is the refund policy?" \
  --out traces/refund_trace.json \
  --verify \
  --report

Evaluate a dataset through the same endpoint when you are ready to regression test:

contexttrace eval \
  --dataset evals/questions.json \
  --endpoint http://localhost:8000/query \
  --method POST \
  --input-key question \
  --answer-path $.answer \
  --contexts-path $.contexts \
  --citations-path $.citations \
  --fail-on "failure_rate>0.25"

Claim-Level Evidence Verification

Verify a portable RAG trace artifact without a hosted dashboard:

contexttrace verify-demo unsupported_claim --report
contexttrace inspect trace.json
contexttrace qa trace.json --corpus docs/ --report
contexttrace verify trace.json
contexttrace verify trace.json --json
contexttrace verify trace.json --report --out reports/example.html
contexttrace verify trace.json --mode semantic
contexttrace verify trace.json --fail-on unsupported --fail-on citation_mismatch
contexttrace verify-benchmark --mode semantic
contexttrace verify-benchmark --mode semantic --report
contexttrace verify-benchmark --case-set external --mode semantic --report
contexttrace compare baseline.json current.json
contexttrace compare baseline.json current.json --report
contexttrace compare baseline.json current.json --fail-on new_failure
contexttrace audit trace.json --corpus docs/
contexttrace audit trace.json --corpus docs/ --report
contexttrace audit trace.json --corpus docs/ --fail-on retrieval_miss
contexttrace audit-benchmark --case-set real --mode semantic
contexttrace audit-benchmark --case-set real --mode semantic --report

Input requires query, answer, and contexts with id and text. Optional citations are checked to catch cited sources that do not actually support the matched claim.

verify-demo uses bundled demo traces, so it works immediately after pip install contexttrace. Available demos include unsupported_claim, partial_support, citation_mismatch, should_abstain, and supported_answer.

Use --mode semantic for local paraphrase-aware matching, and verify-benchmark to inspect bundled precision/recall metrics. The default benchmark includes 32 ContextTrace docs and release-artifact cases. --case-set external adds public OSS documentation and GitHub issue cases from Qdrant, Chroma, Haystack, and LangChain, while --case-set all runs both packs. --report writes an HTML report with misses to inspect.

Verification output includes evidence span offsets, stable span hashes, multiple supporting spans, typed matched/missing facts, and claim-level root causes so partial support failures are easier to inspect.

ContextTrace verifies whether each generated claim is actually supported by retrieved evidence. Instead of only showing a trace or a score, it tells you where the evidence chain broke: unsupported claim, citation mismatch, retrieval miss, answer overreach, conflicting context, or should-have-abstained.

Use the capture helper when you have RAG artifacts in memory:

from contexttrace import capture_rag_trace, write_rag_trace

trace = capture_rag_trace(query=question, answer=answer, contexts=retrieved_docs)
write_rag_trace(trace, "trace.json")

Use contexttrace compare baseline.json current.json to diff two portable traces or saved verify --json outputs. It reports support-rate deltas, new unsupported claims, citation regressions, should-abstain flips, and new root causes, with --fail-on gates for CI.

Use contexttrace audit trace.json --corpus docs/ to diagnose whether an unsupported claim failed because retrieval missed evidence, reranking buried it, chunking omitted the supporting span, the corpus lacks coverage, or generation overclaimed. Audit output includes failure stages, diagnostic signals, and prioritized next actions.

Use contexttrace audit-benchmark --case-set real --mode semantic to test retrieval-audit labels against bundled public OSS documentation and GitHub issue snippets from Qdrant, Chroma, Haystack, LangChain, and ContextTrace docs.

The v0.6.0 verifier uses local lexical heuristics by default. Claim extraction is rule-based, contradiction detection is conservative, and semantic or LLM-judge support can be added later.

What It Catches

• retrieval_miss
• citation_mismatch
• unsupported_answer
• contradicted_answer
• conflicting_sources
• should_have_abstained
• agent failures such as stale_memory_used and tool_error

Privacy

Local mode is the default. ContextTrace makes no network calls unless you configure an LLM judge provider or evaluate a RAG endpoint you provide.

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