Skip to main content
Version: 0.7.2

invera-sdk changelog

All notable changes to the invera-sdk Python package. The Python and TypeScript SDKs share one version and are released together.

[0.7.2] - 2026-11-05​

  • No changes to the Python package's API or behavior. New Hallucination detection recipe on developer.layerwise.ai: a runnable RAG example that streams an answer with stream_with_evidence and shows, flags, or withholds it based on per-claim grounding scores, requires_review, and an optional semantic verifier, plus guidance on calibrating the thresholds on your own data.
  • New LLM judge escalation recipe: extends the hallucination detection recipe so that an LLM judge checks only the claims attribution does not settle, including claims grounded in the question rather than in a retrieved chunk. The judge confirms the hallucination (the claim is withheld) or rejects it (the claim is shown).

[0.7.1] - 2026-11-05​

  • Engines that report contract.chat_completions_endpoint: true or contract.supports_sampling: true are no longer rejected. These flags describe other routes of the Engine server; the SDK never calls a chat route or sends sampling parameters. Compatibility now depends only on how POST /v1/raw_attribution decodes: contract.decoding must be greedy_argmax, otherwise the client raises EngineCompatibilityError before any raw attribution request.
  • The Python SDK now also checks contract.decoding, as the TypeScript SDK already did. EngineContract has a new decoding field (default None, which is rejected); discovery fills it from the Engine's contract metadata.

[0.7.0] - 2026-11-05​

  • Breaking: requires Engine Contract >=1.3.1 <2.0.0 (was >=1.3.0 <2.0.0). Contract 1.3.1 reports sdk_hints.recommended_tool_parser for each model; a Contract 1.3.0 Engine is rejected with EngineCompatibilityError.
  • The client now serves any model its invera Engine reports in GET /v1/models, not only release-pinned profiles. It loads the tokenizer from the reported Hugging Face tokenizer.id and immutable tokenizer.revision and accepts it only when its locally computed fingerprint equals tokenizer.fingerprint. A new model or tokenizer revision needs no SDK release.
  • model_type is no longer an allowlist. Discovery is instead checked for a pinned tokenizer identity (40-character commit revision, SHA-256 fingerprint), an available chat template, valid EOS and special token ids, and a supported sdk_hints.recommended_tool_parser. Each failure raises EngineCompatibilityError before any POST /v1/raw_attribution; a tokenizer fingerprint mismatch now raises EngineCompatibilityError instead of RuntimeError.
  • Models whose Engine recommends no tool parser stream text completions; a request with tools for such a model raises EngineCompatibilityError.
  • Tool calls are now parsed in both formats the Engine recommends: qwen_json_tool_call, Hermes-style JSON inside <tool_call> (Qwen2.5, Qwen3, QwQ), and qwen_xml_tool_call, XML per parameter (Qwen3.5 and later). The client picks the format from sdk_hints.recommended_tool_parser. Streaming chunks, ParameterEvidence, and the pre-execution tool-call decision are equivalent for the same call in either format; a malformed JSON call is reported with the schema issue code invalid_json.
  • Breaking (advanced): parse_qwen_tool_calls and stream_evidence require the tool parser (stream_evidence(..., tool_parser=...)); there is no implicit qwen_xml_tool_call default. Pass the Engine's recommended_tool_parser; stream_evidence accepts None for streams without tool calls.
  • Discovery must report exactly one model per Engine endpoint; zero or several entries raise EngineCompatibilityError (several were previously resolved to the first supported entry). The SDK still never asks the Engine for a specific model.
  • A custom TokenizerAdapter passed to create_client_with_tokenizer now works for any model whose identity exactly matches discovery, without static profile resolution.
  • Breaking (invera_sdk.advanced): MODEL_PROFILES, ModelProfile, and SUPPORTED_MODEL_PROFILES are removed; the new SUPPORTED_TOOL_PARSERS lists the parsable tool-parser hints. create_hugging_face_tokenizer now takes the expected TokenizerIdentity instead of a ModelProfile.

[0.6.1] - 2026-10-02​

  • Supports invera Engines serving Qwen/Qwen3.8-27B through the new release-pinned qwen3.8-27b tokenizer profile. The client still picks the profile from the tokenizer identity the Engine reports; there is no model option.

[0.6.0] - 2026-10-02​

  • Breaking: the SDK now supports only invera Engines serving Qwen/Qwen3.5-4B (tokenizer profile qwen3.5-4b, model family qwen3_5). Qwen/Qwen3-4B-Instruct-2507 is no longer supported.
  • Breaking: the model client option is removed. The client always uses the release-pinned model profile; it cannot choose a model.
  • Breaking: requires Engine Contract >=1.3.0 <2.0.0 (was >=1.2.0 <2.0.0).
  • Breaking: tool calls are parsed only in the Qwen3.5 XML format (<function=…><parameter=…>…</parameter></function>); parameter values are typed from the tool's JSON schema. A malformed call is reported with the new schema issue code invalid_xml, which replaces invalid_json.
  • The default tokenizer adapter now works with Transformers 4.x as well: its tokenizer fingerprint no longer depends on the installed Transformers major version, which previously failed the pinned-profile check on 4.x.
  • tool-description messages are display-only: the default tokenizer never renders them into the prompt. Pass tool definitions through tools.

[0.5.2] - 2026-09-25​

  • New api_key client option: a static Engine API key sent as Authorization: Bearer <key> on every Engine request.

[0.5.0] - 2026-09-22​

  • First release published to PyPI.
  • Licensed under the PolyForm Perimeter License 1.0.1 together with the Layerwise SDK Additional Permission; both texts ship in the distribution.
  • The package README is reduced to name, version, vendor, documentation link, license pointer, and this changelog.

[0.4.3] - 2026-09-21​

  • No Python changes; version kept in lockstep with the TypeScript SDK.

[0.4.2] - 2026-09-21​

  • invera_sdk.__version__ added as an alias of SDK_VERSION.

[0.4.1] - 2026-09-17​

  • Output words are split at internal whitespace in attribution word mapping.

[0.4.0] - 2026-09-02​

  • Emitted attribution word mapping aligned between Python and TypeScript.
  • Releases are gated by live Engine end-to-end checks.

[0.3.2] - 2026-08-25​

  • Supports invera Engine Contract 2.x on the v1 routes.

[0.3.1] - 2026-07-30​

  • Renamed to invera. Golden Path client, evidence stream, and the invera_sdk.advanced extension module.