Instructor Python Files

Added

• Bedrock native structured outputs: Add explicit Mode.JSON_SCHEMA and Mode.TOOLS_STRICT support through Converse outputConfig.textFormat and strict tool schemas, with recursive schema normalization and a boto3 1.42.42 minimum. Model selection remains caller-controlled. (#2515, #2084, #2086)
• Validation retry budgets: Add positive cumulative token_budget limits for structured non-streaming retries, immutable completion:usage snapshots, sync/async cutoff parity, and stable cumulative usage metadata. Valid responses still win after crossing the budget. When a budget is configured, a failed attempt with unavailable usage stops before another provider call. (#2512, #2391, #2392)

Changed

• Contributor workflow: Align contributor setup and dependency management around locked uv sync environments, uv run commands, and uv add only for intentional project metadata changes. (#2516, #2354)

Fixed

• Mistral SDK compatibility: Support the mistralai 2.x client export on Python 3.10+ while retaining the compatible 1.x fallback required by Python 3.9. (#2513, #2298, #2365)
• Bedrock reasoning JSON: Parse the final complete JSON value after reasoning text or <think> blocks, preserve JSON escape sequences, and keep caller-owned messages unchanged during Bedrock request preparation and retries. (#2514, #2076, #2287)
• Remote multimodal fetches: Apply the existing 30-second request timeout to image, audio, and PDF downloads so an unresponsive URL cannot block a caller indefinitely. (#2507)
• OpenAI streaming retries: Keep TOOLS, JSON, JSON_SCHEMA, and MD_JSON retries on the streaming parser after the one-shot model marker is consumed, allowing corrected streamed responses to validate successfully. (#2508)
• Iterable streaming unions: Parse PEP 604 unions (create_iterable(response_model=Weather | GoogleSearch), Iterable[Weather | GoogleSearch]) member by member instead of calling model_validate_json on types.UnionType. (#2509)
• Package metadata: Point the published distribution's repository URL at the current 567-labs/instructor organization and validate it before release.
• Retry usage accounting: Accumulate nested and newly added numeric usage fields across OpenAI and Anthropic retries, including prediction, cache-write, cache-creation, and server-tool counters, without treating boolean metadata as billable usage. (#2493, #2500)
• OpenAI Responses reask: Add a fallback correction message when a RESPONSES_TOOLS response contains no tool calls (e.g. reasoning-only output), so retries carry validation feedback instead of resending the identical request. (#2498)
• v2 parallel tools: Preserve raw iterable type hints through the sync and async patch wrappers so parallel tool schemas and results retain every requested model type. (#2501)
• Credential redaction: Hide common OAuth and Google API credential aliases in nested v2 debug logging while preserving non-secret token configuration. (#2490, #2491)
• Retry and message integrity: Preserve cache keys and caller-owned retry messages, retain empty-content legacy function calls, return Anthropic tool results for every parallel tool call, and handle missing OpenAI/Mistral tool calls as retryable parse failures. (#2454, #2455, #2464, #2484, #2485, #2486, #2448, #2453)
• Streaming and DSL correctness: Isolate partial-model recursion guards, preserve partial nested models and explicit nulls, harden citation matching, derive useful Iterable union names, and continue scanning JSON streams after non-JSON or multiple balanced values. (#2422, #2430, #2431, #2452, #2456, #2461, #2463, #2476, #2487, #2489)
• Provider request handling: Avoid mutating Gemini generation config and cached OpenAI schemas, disable Anthropic parallel calls for forced single-tool requests, forward Bedrock default models, and label OpenAI audio as WAV or MP3 without misrepresenting unsupported formats. (#2450, #2451, #2465, #2467, #2477, #2478, #2447, #2415)
• Batch, CLI, and citation runtime: Accept valid empty batch objects, use typed OpenAI file attributes in the CLI, normalize None message content, and install regex as the direct dependency required by CitationMixin. (#2473, #2441, #2440, #2443)
• Provider documentation: Refresh retired Cerebras model IDs, clarify current and deprecated Google provider prefixes, and fix the Vertex Google GenAI example so its default model is passed to from_genai(). (#2494, #2289, #2343, #2416, #2475)
• Multimodal (Audio): Raise explicit ValueError or FileNotFoundError from Audio.from_url() and Audio.from_path() instead of relying on bare assert statements that can disappear under python -O. (#2361)
• v2 message handling: Preserve caller-owned message lists and nested content across request preparation and retries for OpenAI-compatible, Cohere, Mistral, OpenRouter, Writer, and xAI handlers. (#2417, #2428)
• v2 JSON extraction: Prefer the final complete top-level JSON value in text responses and retain every JSON object when multiple objects arrive in one streaming chunk.
• v2 schemas: Treat fields with Pydantic default_factory values as optional in generated OpenAI tool schemas.
• v2 partial streaming: Build model instances for present Optional[BaseModel] fields during incomplete streams instead of exposing raw dictionaries.
• v2 iterable unions: Generate stable member-derived names such as IterableAOrB for both Union[A, B] and A | B response models.
• Mistral/Vertex AI partial streaming: Avoid forwarding iterable-only parser arguments into completed Partial responses, preventing final Pydantic validation errors for sync and async streams.
• Batch providers: Handle missing optional SDKs safely, validate OpenAI batch input before client setup, report exhausted output-file retries clearly, and remove unreachable fallbacks.
• OpenAI/Writer tools: Raise clear response-parsing errors for completions with no choices or tool calls instead of leaking attribute and index errors.
• Fireworks streaming: Keep non-streaming async calls non-streaming and return streaming async generators without incorrectly awaiting them.
• GenAI uploads: Respect max_retries=0 without an unwanted sleep or polling request, allow recovery on the final permitted retry, and report nameless pending uploads clearly before polling.
• Gemini/GenAI messages: Honor an explicit system message for unstructured requests, remove the unsupported raw system argument, and reject invalid scalar message content clearly.
• Templating: Use populated contents when messages is empty, avoid mutating nested caller input, and preserve uncopyable metadata during template expansion.
• Anthropic system messages: Reject invalid new system-message values even when no existing system message is present.
• Python 3.9: Include the required type-evaluation backport in minimal installs, keep overload metadata available, and avoid runtime evaluation of unsupported union syntax in the core response path and offline tests.
• v2 imports: Defer OpenAI SDK imports from core v2 modules until an OpenAI-specific path actually needs them, reducing import side effects for non-OpenAI usage. (#2390)
• v2 response models: Treat list[A | B] PEP 604 unions of Pydantic models as iterable response models, matching list[Union[A, B]] schema behavior. (#2377)
• OpenAI Responses API: Align RESPONSES_TOOLS text.format with the forced tool schema and add targeted retry guidance when tool calls return empty {} arguments. (#2300, #2304)

Security

• LLM validator isolation: Send validation rules and candidate values as structured JSON data under a fixed trusted instruction to reduce prompt-injection risk, and raise ValueError for rejected values instead of relying on optimization-sensitive assertions. (#2511, #2056, #2307)

Tests / CI

• Fork-safe contributor checks: Mark auto-client network tests explicitly and exclude them from core, coverage, and release lanes so fork PRs without provider secrets do not fail with empty authorization headers.
• Coverage and test quality: Run the complete offline suite on Python 3.9-3.13, enforce fork-safe statement and branch coverage plus supported-version type checks in pull-request CI, add strict resource and thread warning checks, and provide a manual retry-mutation workflow. Consolidate typed response, stream, and SDK fixtures; remove duplicate tests and unreachable provider paths; and replace coverage-only stubs with meaningful edge-case and transport-backed provider checks.
• Release safety: Validate the declared source, lockfile, changelog, tag, and built artifacts before any publication step; require an explicit version confirmation and publish opt-in; and publish the exact tested assets instead of rebuilding from a moving branch.