AgentClaimGuard v0.4.1 - RAGFlow evidence example and integration patterns

AgentClaimGuard v0.4.1

This is a documentation and integration example maintenance release.

Added

• Added a RAGFlow-style evidence provider example.
• Added copyable retrieved chunk and mapped evidence JSON examples.
• Added a runnable examples/ragflow_evidence/demo.py example.
• Added platform integration pattern documentation for:
  • HTTP tool integrations
  • Evidence provider integrations
  • Framework adapter integrations

Changed

• Clarified that RAGFlow-style integrations should map retrieved chunks into AgentClaimGuard Evidence records.
• Clarified that retrieval, ranking, and vector search stay outside AgentClaimGuard.
• Updated README, adapter docs, example docs, and changelog.

Validation

python -m pytest -q
python -m compileall agentclaimguard examples tests
python examples/ragflow_evidence/demo.py
python examples/claim_extraction/demo.py
python examples/langchain_guard/demo.py
python examples/langgraph_guard/demo.py
python -m build
python -m twine check dist/*

Expected result:

33 passed
compileall passed
RAGFlow evidence demo passed
claim extraction demo passed
LangChain demo passed
LangGraph demo passed
build passed
twine check passed

Scope

This release does not add a RAGFlow plugin package, vector search, ranking logic, or verifier/runtime changes.

AgentClaimGuard v0.4.0 - Optional claim extraction helpers

AgentClaimGuard v0.4.0

This release adds optional deterministic claim extraction helpers.

The new helpers make it easier to turn user-provided or LLM-generated claim candidates into structured AgentClaimGuard Claim objects.

Important boundary:

Extraction != Verification

Added

• Added agentclaimguard.extractors
• Added ClaimCandidate
• Added ClaimExtractionResult
• Added ClaimExtractionTemplate
• Added create_claim_from_text(...)
• Added create_claims_from_items(...)
• Added claim extraction demo
• Added claim extraction documentation and tests

Behavior

The extraction helpers:

• Do not call any LLM
• Do not verify truth
• Do not perform evidence matching
• Do not change verifier semantics
• Only help structure claim-like inputs into AgentClaimGuard claims

Validation

python -m pytest -q
python -m compileall agentclaimguard examples tests
python examples/claim_extraction/demo.py
python examples/langchain_guard/demo.py
python examples/langgraph_guard/demo.py
python -m build
python -m twine check dist/*

Expected result:

33 passed
compileall passed
claim extraction demo passed
LangChain demo passed
LangGraph demo passed
build passed
twine check passed

A local OpenAI-compatible LLM smoke test also validated the intended optional pattern:

LLM output -> claim candidates -> create_claims_from_items(...) -> AgentClaimGuard.verify(...)

The smoke test did not commit any endpoint, API key, temporary script, or local LLM configuration.

AgentClaimGuard v0.3.2 - Publishing workflow and Dify HTTP example

AgentClaimGuard v0.3.2

This is a documentation, publishing workflow, and integration example release.

Added

• Added a manual GitHub Actions workflow for PyPI Trusted Publishing.
• Added a Dify HTTP tool integration example using the existing FastAPI /v1/verify endpoint.
• Added copyable request and blocked-response payload examples for Dify-style HTTP tool workflows.

Changed

• Updated README quickstart to use the published PyPI package.
• Updated release checklist to prefer PyPI Trusted Publishing.
• Documented manual token upload as a fallback path.
• Expanded adapter and example documentation for the Dify HTTP tool flow.

Validation

python -m pytest -q
python -m compileall agentclaimguard examples tests
python examples/langchain_guard/demo.py
python examples/langgraph_guard/demo.py
python -m build
python -m twine check dist/*

Expected result:

23 passed
compileall passed
LangChain demo passed
LangGraph demo passed
build passed
twine check passed

Publishing

This release is intended to be the first real validation of the PyPI Trusted Publishing workflow.

AgentClaimGuard v0.3.1 - Packaging and metadata maintenance

AgentClaimGuard v0.3.1

This is a packaging and metadata maintenance release.

Maintenance

• Restored the canonical Apache-2.0 license text for GitHub license detection.
• Added explicit copyright ownership for Hao Peng in project metadata.
• Added PyPI project URLs.
• Added release tooling optional dependencies for build and twine.
• Updated installation instructions for PyPI-based installs.
• Updated the release checklist with build, twine check, TestPyPI, and fresh virtual environment validation.
• Removed duplicate policy-file wheel packaging configuration while preserving built-in policy files.

Validation

python -m pytest -q
python -m compileall agentclaimguard examples tests
python examples/langchain_guard/demo.py
python examples/langgraph_guard/demo.py
python -m build
python -m twine check dist/*

Expected result:

23 passed
compileall passed
build passed
twine check passed

TestPyPI dry run was also validated before this release.

AgentClaimGuard v0.3.0 - LangChain Runnable adapter

AgentClaimGuard v0.3.0

This release adds the LangChain Runnable adapter for AgentClaimGuard.

AgentClaimGuard can now wrap LangChain Runnable chains and attach claim-level verification results to their outputs.

Added

• Added agentclaimguard.adapters.langchain
• Added create_guarded_runnable(...)
• Added LangChain Runnable wrapper support
• Added sync invoke(...) and async ainvoke(...) support
• Added custom field_map support
• Added result_key customization
• Added optional overwrite_result=True behavior for explicit result-key overwrites
• Added LangChain Runnable demo
• Added LangChain adapter tests
• Added LangChain adapter documentation and README examples

Behavior

The LangChain adapter:

• Runs the underlying Runnable first
• Resolves claims, evidence, and tool_results from Runnable output first
• Falls back to Runnable input when output fields are missing
• Supports callable field extractors
• Requires resolved fields to be lists or None
• Attaches guard_result to the output
• Raises ValueError on result-key collisions unless overwrite_result=True

Example output

sync_final_answer=Revenue increased by 15%.
sync_guard_status=blocked
sync_claim_status=tool_required
async_final_answer=Revenue increased by 15%.
async_guard_status=blocked
async_claim_status=tool_required

Validation

python -m pytest -q
python -m compileall agentclaimguard examples tests
python examples/langchain_guard/demo.py
python examples/langgraph_guard/demo.py

Expected result:

23 passed
compileall passed
LangChain demo passed
LangGraph demo passed

Scope

This release intentionally only adds the LangChain Runnable wrapper.

It does not add LangChain middleware, automatic claim extraction, Dify integration, RAGFlow integration, or LLM-as-verifier support.

AgentClaimGuard v0.2.1 - LangGraph adapter polish and maintenance

AgentClaimGuard v0.2.1

This is a maintenance and documentation polish release for the LangGraph adapter line.

Added

• Added clearer LangGraph demo notes under examples/langgraph_guard/.
• Added troubleshooting notes for common LangGraph adapter issues.
• Added a release checklist for maintainers.
• Added a pull request template for release-oriented validation.

Changed

• Clarified LangGraph adapter usage in README and docs/adapters.md.
• Documented typed state usage for preserving guard_result.
• Documented full START / END / repair / human_review wiring.
• Clarified that custom result_key must stay consistent between the guard node and router.
• Improved issue templates for integration-focused reports.
• Bumped package and FastAPI server versions to 0.2.1.

Fixed

• Fixed the LangGraph demo fallback path so it preserves guard_result before updating final_answer.

Maintenance

• Updated GitHub Actions workflow dependencies to Node 24 compatible versions.

Validation

python -m pytest -q
python -m compileall agentclaimguard examples tests
python examples/langgraph_guard/demo.py

Expected result:

13 passed
compileall passed
guard_status=blocked
claim_status=tool_required
route=blocked

AgentClaimGuard v0.2.0 — LangGraph node adapter

AgentClaimGuard v0.2.0

This release adds the first framework adapter for AgentClaimGuard: a LangGraph node adapter.

AgentClaimGuard can now be embedded directly into LangGraph workflows as an evidence guard node for LLM agent claims.

Added

• agentclaimguard.adapters.langgraph
• create_evidence_guard_node(policy=...)
• route_by_guard_status(state, result_key="guard_result")
• Policy.load_builtin(...)
• LangGraph adapter tests
• Runnable LangGraph guard demo
• Adapter documentation and README updates

Example

The new demo shows an agent-generated numeric claim being blocked because it lacks the required calculator tool result:

guard_status=blocked
claim_status=tool_required
route=blocked

Validation

python -m pytest -q
python -m compileall agentclaimguard examples tests
python examples/langgraph_guard/demo.py

Expected result:

12 passed
compileall passed
guard_status=blocked
claim_status=tool_required
route=blocked

AgentClaimGuard v0.1.1 ? Project homepage and contributor workflow polish

AgentClaimGuard v0.1.1 is a documentation and project polish release.

Added

• README badges
• Shorter quickstart
• Example output documentation
• Clearer roadmap
• Adapter integration notes
• GitHub issue templates

Changed

• Improved project homepage positioning
• Clarified contributor workflow

Maintenance

• Added .playwright-mcp/ to .gitignore
• Verified tests and compile checks

AgentClaimGuard v0.1.0 ? Initial claim-level evidence gate runtime

AgentClaimGuard v0.1.0

This is the first runnable open-source release of AgentClaimGuard.

AgentClaimGuard is a framework-agnostic evidence gate for LLM agent claims. It verifies whether important claims in LLM / RAG / Agent outputs are supported by evidence, tool results, and user-defined policies.

Core idea:

Claim -> Evidence -> Tool -> Verify

Principles:

No evidence, no claim.
No tool result, no numeric conclusion.
No source, no compliance judgment.

What is included

• Python SDK
• Pydantic schemas for:
  • claims
  • evidence
  • tool results
  • policies
  • verification results
• YAML-based policy runtime
• Core verifier
• Basic validators:
  • required evidence check
  • required tool result check
  • citation binding check
  • forbidden verdict check
  • conflicting evidence check
• Built-in generic policies
• FastAPI server with:
  • GET /health
  • POST /v1/verify
  • POST /v1/repair
• Three runnable demos:
  • numeric conclusion
  • compliance judgment
  • RAG citation
• Initial documentation
• GitHub Actions CI
• Apache-2.0 license
• Pytest test suite

What AgentClaimGuard is not

AgentClaimGuard is not an Agent framework, RAG engine, vector database, structured output library, or general-purpose safety guardrail.

It is a lightweight claim-level reliability layer that can be embedded into existing LLM, RAG, and Agent workflows.

Validation

This release has been validated with:

python -m pytest -q
python -m compileall agentclaimguard examples tests

Expected result:

8 passed
compileall passed

The three demos also produce the expected blocked / need_check / insufficient_evidence results.

Roadmap

Next planned directions:

• LangGraph node adapter
• LangChain middleware / Runnable wrapper
• OpenAPI integration examples
• Optional claim extraction helper
• Dify / RAGFlow integration examples
• More generic policy templates