Lár v1.7.0 Release Notes

Lár v1.7.0 Release Notes

What's New

1. Enterprise Streaming & Structured JSON Capabilities (LLMNode)

LLMNode has received a major upgrade with enterprise-grade features seamlessly integrated into the LiteLLM backend.

• Streaming Support: Direct terminal streaming capability. When instantiated with stream=True, output is streamed chunk-by-chunk to sys.stdout while generating a cohesive total_tokens usage metric under the hood.
• Structured Output Support: Easily ensure outputs match Pydantic schemas by passing response_format=<PydanticModel>.
• Snath Cloud Hooks: New parameters fallbacks, caching, and success_callbacks have been added to prepare for deep infrastructure observability hooks and integrations.

2. Precise Token Budget Merging (BatchNode)

• Thread-Safe Budget Deductions: State modifications executed in parallel across BatchNode threads now strictly compute delta token costs to avoid race-condition overwriting. The parent graph mathematically merges total budgets correctly from all worker nodes.

3. Compliance and Audit Trails (For Auditors)

• We have introduced a formal Verification Script (examples/compliance/11_verify_audit_log.py) directly for Compliance officers and auditors to authenticate JSON audit log files offline via HMAC-SHA256 signatures to certify anti-tampering (Crucial for FDA/EU-AI Act adherence). Usage documentation has also been promoted to the README.md.

Bug Fixes & Refactors

• Updated standard unit tests to correctly mock API generations for streaming branches and to ensure budget accuracy upon branch reconciliation.