New Feature: Python — provider-independent batch request API for model services (encode/create/status/cancel/collect)

未关闭
#14,746 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 3 天内回复

还没有人认领这个 Issue。

评估

难度
5/5
预计耗时
一周以上
新手友好度
35/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
活跃
技术栈
python

调研方向

Read the proposed batch API scope and compare the referenced pydantic-ai batch codec with Semantic Kernel’s Python connector structure; the issue does not name specific SK Python files. Start by discussing the API design with maintainers, as the contribution policy requires agreement before implementation. Done means an agreed design, then a provider implementation with tests for credential grouping, polling, errors, and large embedding batches.

由索引模型根据 Issue 内容生成。

描述

.NET python triage

Semantic Kernel currently invokes one AI service per call via service IDs, and large-input workloads are handled ad hoc inside connectors (e.g. the 100-input Gemini embedding chunking in PR #14700). Peer framework pydantic-ai provides a shared batch model: encode requests, group them by credential, create/status/cancel provider batch jobs, and report per-request results including unrun requests.

Proposal: add an opt-in batch request API for SK Python connectors (BatchRequest/Batch/result types plus batch_encode/create/status/cancel on a model-service interface), implemented first for one provider, so future connectors get consistent grouping, polling and error behavior. No public signatures change for existing non-batch paths; this needs design agreement before implementation per the contribution policy.

Users

SK users indexing document sets or running large offline evaluation/scoring workloads who today must chunk inputs and poll provider batch jobs themselves, per connector.

Expected Effect

A shared batch abstraction lets connectors (OpenAI, Azure OpenAI, Google, Anthropic) support provider batch endpoints with consistent grouping, polling and error semantics instead of ad-hoc per-connector chunking like PR #14700.

Scope

Python first (where connector activity is concentrated): a new batch module under semantic_kernel/connectors or services plus opt-in implementation on one connector (e.g. OpenAI chat or Google embeddings). .NET parity and additional providers stay out of scope until the design is agreed.

Validation

cd python && uv run pytest tests/unit/connectors -q plus new tests: encode produces provider payloads; create groups only same-credential requests; status polling stops at ended; unrun/canceled requests surface typed errors; >100-input embedding calls split like #14700. Expected: new tests fail on base, pass with change.

Risks

New public API surface in a compatibility-bound project; provider batch semantics differ (Google inline vs OpenAI file-based jobs), so a wrong abstraction could force breaking changes later.

Effort

Design discussion plus ~1-2 weeks: shared batch types and codec-equivalent helpers, one connector implementation, unit tests mirroring pydantic-ai's grouped/polling/error coverage.

Comparison

pydantic-ai implements this as provider-independent codecs shared by anthropic/_batch.py, _openai_batch.py and _google_batch.py with BatchJobError, group isolation by credential hash and byte-limit prechecks; SK has only ordered service selection (MultipleModelTests) and per-connector workarounds.

Maintainer direction requested before implementation.

AI-assisted contribution.

主要语言
C#
星标
28.8k
派生
4.9k
平均合并
2 天 17 小时
30 天内合并 PR
48

环境准备

在 Codespaces 中打开

在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

microsoft/semantic-kernel 的其他 Issue

查看 microsoft/semantic-kernel 的全部 Issue

相似的 Issue

更多 C# Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。