latence 0.1.3

What's Changed

• Fixes SDK pipeline uploads for multi-MB local files.
• client.pipeline.run(files=[...]) now preserves the presigned B2 upload path for larger files instead of sending multi-MB base64 JSON through the gateway.
• Avoids Cloudflare Worker resource-limit failures before pipeline jobs are created.

Validation

• python -m pytest tests/test_refinery.py::TestPipelineRun -q
• python -m build

v0.1.2 — Trace service + README restructure

Highlights

New: client.experimental.trace

First-class wrapper for the Latence Trace service — groundedness, phantom-hallucination, and drift scoring over one pod, three lanes:

• client.experimental.trace.rag(response_text=..., raw_context=..., ...) — score a RAG answer for groundedness, context coverage, and unused context.
• client.experimental.trace.code(response_text=..., raw_context=..., response_language_hint=..., session_state=...) — agentic-coding turn scoring with opaque cross-turn session chaining.
• client.experimental.trace.rollup(turns=[...]) — stateless, sub-ms aggregation of N per-turn outputs into one session scoreboard.

Full sync + async parity (AsyncLatence), background-job submission (return_job=True), SupportUnitInput structured premises, and strict client-side validation.

Fixed: production-response schema drift

• TraceResponse.primary_metric is now str | None (was Literal[...]) — the pod may echo derived metric names like groundedness_v2.
• TraceRollupResponse.drift_trend is now dict[str, Any] | None (was list[float]) — matches the live {last, max, mean, min, ...} payload shape.

Docs

• README.md restructured into a three-step story: Trace your answers → Upgrade data quality → Upgrade retrieval, with a 20-second happy path and signal-to-action routing.
• docs/trace.md is the polished reference for every param, response field, and ValueError.
• New portal docs hub at https://latence.ai/docs/trace.

Other

• experimental.trace exposed alongside existing experimental.* services.
• 5 advanced RAG/code kwargs added for full gateway parity: raw_context_chunk_tokens, response_chunk_tokens, query_prompt_name, document_prompt_name, debug_dense_matrices.
• Regression tests against real production responses.

Install

pip install --upgrade latence

v0.1.1

Latence Python SDK v0.1.1