任务模型：嵌入、重排、OCR 与语音识别Task models: embedding, reranking, OCR and ASR从业者视角的方法论 · 关键事实挂官方出处practitioner-perspective methodology · key facts carry official sources

聊天生成、嵌入、重排、文档识别、语音识别是不同接口。先写输入类型、输出结构和验收方法，再查模型。下列四个模型卡只是接口例子，不是排名、全量名单或部署推荐；没有在本站下载或运行权重。核对范围为 2026-11-04 的官方模型卡示例与限制。

不熟悉输入输出概念时先读 语音识别，区分音频转写和文字生成，再回这里选任务接口。

平台支持按模型版本、库版本、硬件后端和任务一起核对。提供 Python 示例不等于已经确认 Windows、macOS 或所有 Linux GPU 可用；本文没有安装这些环境，未确认的平台明确保留未知。

Chat generation, embedding, reranking, document recognition and speech recognition expose different interfaces. Define input, output and acceptance checks before choosing a model. The four cards below illustrate contracts; they are not a ranking, exhaustive catalogue or deployment recommendation. No weights were downloaded or executed here. The 2026-11-04 check covers official model-card examples and limitations.

Check platform support by model version, library version, hardware backend and task together. A Python example does not establish Windows, macOS or every Linux GPU combination. Those environments were not installed here; unverified platforms remain unknown.

For input/output concepts, start with speech recognition to distinguish audio transcription from text generation, then return to the task contract.

接口例子：BAAI/bge-m3。模型卡展示文本编码与多种检索表示；本节仅讨论文字检索，不推断它能直接读音频或扫描页。

• 输入：查询和文档文字，先保留分块 ID、版本与出处。
• 输出：所选模式的表示；向量不是自然语言答案。记录维度、归一化、查询/文档处理方式与模型修订。
• 运行说明：按模型卡中的 FlagEmbedding 或所列库示例核对依赖和调用；不同模式不假设互换。
• 验收：固定一组有出处的查询与相关段落，检查召回和未命中；不要只看向量维度或榜单。换模型后检查是否需要重建索引，不能把两个模型的向量混入旧索引。
• 未确认：本文未验证具体操作系统、CPU/GPU 后端、量化格式或长输入实际占用。

继续读 怎么挑 embedding 模型；召回有材料但答案不对时，分别检查检索与生成。

Interface example: BAAI/bge-m3. Its card shows text encoding and multiple retrieval representations. This section concerns text retrieval, not direct audio or scan input.

• Input: query/document text with chunk IDs, version and source.
• Output: representations for the chosen mode, not a natural-language answer. Record dimensions, normalization, query/document processing and model revision.
• Runtime: follow the card’s FlagEmbedding or listed library examples; do not assume modes are interchangeable.
• Acceptance: use fixed queries and located relevant passages to inspect recall and misses. Dimensions or leaderboard position alone are insufficient. Check reindexing requirements when changing models; never mix incompatible embeddings into the old index.
• Unknown here: exact OS, CPU/GPU backend, quantization format and real long-input allocation.

Continue with choosing embeddings. Check retrieval and generation separately when a relevant source is found but the answer is wrong.

接口例子：BAAI/bge-reranker-v2-m3。官方示例用查询和候选文本对计算分数；不是先把整个库重新变成嵌入。

• 输入：同一查询与已召回的候选段落，每条保留候选 ID 和原顺序。
• 输出：候选分数与重排顺序；分数不是事实正确率，也不能跨不同模型直接当同一刻度。
• 运行说明：模型卡列出 FlagReranker 与 Transformers 等用法；确认输入成对、截断规则和输出分数解释。
• 验收：先判断正确段落是否已进入候选集；没有召回的段落无法靠这次重排补回。对照有判定依据的排序，分别记录相关性与答案证据。
• 未确认：本文未验证各平台后端、服务并发或与某个本地聊天运行器的兼容性。

读 怎么给检索结果重排。少量候选已足够时可不加重排，不把它写成 RAG 必经步骤。

Interface example: BAAI/bge-reranker-v2-m3. Official examples score query/candidate text pairs, rather than rebuilding corpus embeddings.

• Input: one query paired with retrieved passages, retaining candidate IDs and original order.
• Output: scores and reordered candidates. Scores are not factual accuracy, and different models do not share a guaranteed scale.
• Runtime: the card lists FlagReranker and Transformers usage. Check paired input, truncation and score interpretation.
• Acceptance: first establish whether the right passage reached the candidate set; reranking cannot recover an absent candidate in that pass. Compare evidence-backed relevance judgements, recording answer support separately.
• Unknown here: platform backends, serving concurrency and compatibility with a particular local chat runner.

Read reranking retrieved results. Reranking is optional when existing candidates suffice; it is not a mandatory RAG stage.

接口例子：PaddlePaddle/PaddleOCR-VL。模型卡提供文档图像处理及保存 JSON/Markdown 的管线示例。文档解析管线与单次模型调用不是同一层，不能只凭权重格式判断完整流程可用。

• 输入：原页图像及需要的解析任务；PDF 有可靠文字层时先评估直接抽取。
• 输出：文本、结构或版面结果，具体字段按所用管线版本核对；不是自动核实过的业务表。
• 运行说明：从卡片链接的 PaddleOCR-VL 安装文档核对管线、依赖、后端和版本；区分本卡与其他后续版本。
• 验收：保留页码/区域，检查阅读顺序、表头、跨行内容、空格、前导零和数字；金额另复算。
• 未确认：本文未安装管线，没有验证 Windows/macOS 支持、特定显卡、GGUF 版本或扫描准确率。

用 虚构 PDF 转表格任务 练习，参考 CSV 是设计答案，不能标成该模型输出。

Interface example: PaddlePaddle/PaddleOCR-VL. The card provides a document-image pipeline with JSON/Markdown export. A parsing pipeline and one model call are different layers; a weight format alone does not establish the complete workflow.

• Input: original page images and the parsing task; evaluate direct extraction when a PDF has usable text.
• Output: text, structure or layout according to the selected pipeline version, not a verified business table.
• Runtime: follow linked PaddleOCR-VL installation guidance for pipeline, dependencies, backend and versions. Distinguish this card from subsequent releases.
• Acceptance: retain page/region evidence and check reading order, headers, wrapped rows, blanks, leading zeros and numbers. Recalculate money separately.
• Unknown here: Windows/macOS support, exact GPU, GGUF availability and scan accuracy; no pipeline was installed.

Practice with the fictional PDF-to-table task. Its designed reference CSV must not be attributed to this model.

接口例子：openai/whisper-large-v3。模型卡提供语音识别和翻译说明及 Transformers 示例，并列出转写局限。

• 输入：音频、语言/任务设置；解码与重采样由所选管线要求决定，不能只改文件后缀。
• 输出：转写或指定的翻译结果；时间戳配置需单独核对。转写不自动等于说话人分离、会议决定或行动项。
• 运行说明：按模型卡核对处理器、模型加载与音频管线。本文没有确认任意第三方客户端或量化仓可直接替代。
• 验收：回听姓名、数字、否定、静音和条件；生成式转写可能加入没有说出的文字。纪要再查决定/提议和负责人。
• 未确认：本文未运行 ASR、分离多人或验证具体平台上的延迟/内存。

用 合成会议录音任务 练习，输入只有一个声音朗读角色标签，不能用于宣称多人分离效果。

Interface example: openai/whisper-large-v3. The card describes recognition/translation, Transformers usage and transcription limitations.

• Input: audio plus language/task settings. Follow pipeline decoding/resampling requirements; renaming an extension does not convert audio.
• Output: transcription or requested translation; check timestamp settings separately. Transcription does not automatically supply speaker diarization, meeting decisions or actions.
• Runtime: verify processor, loading and audio pipeline against the card. No arbitrary third-party client or quantization repo is established as a replacement here.
• Acceptance: replay names, numbers, negation, silence and conditions; generative transcription can add unspoken text. Check proposals/decisions and owners in notes separately.
• Unknown here: ASR runs, multi-speaker separation and platform-specific latency/memory were not tested.

Try the synthetic-meeting task. One voice reads role labels; it cannot establish diarization quality.

给每个候选建立一条记录：模型仓与固定 revision；任务入口；输入预处理器/tokenizer；输出格式；权重格式与精度；运行库/运行器版本；操作系统；CPU/GPU/驱动后端；所需自定义代码；许可来源；已核对范围与未确认项。

• safetensors 是权重存储格式，不能据此推断模型架构、任务接口和硬件支持。
• GGUF 文件存在不等于任意 GGUF 运行器支持该架构、处理器与任务。先查运行器对该模型/任务/版本的明确说明；未知就保留未知。
• OpenAI 风格接口或相同文件扩展名不是行为一致的证据；嵌入、评分、识别和聊天的请求/响应分别检查。
• 需要执行仓库自定义代码时，把代码与版本作为独立依赖审阅；不要为了消除报错无条件打开远程代码执行。

上面的模型卡有不同库示例，本文不提供一张猜测出来的“所有平台都支持”表。没有安装验证的组合不能写“已兼容”。

For each candidate record model repository and pinned revision, task entry point, preprocessor/tokenizer, output format, weight format/precision, library/runtime version, OS, CPU/GPU/driver backend, required custom code, licence source, check scope and unknowns.

• safetensors stores weights; its presence alone does not establish architecture, task contract or hardware support.
• A GGUF file does not establish support in every GGUF runner for its architecture, processor and task. Seek an explicit model/task/version statement; retain unknowns otherwise.
• An OpenAI-style interface or shared extension is not behavioral equivalence. Check embedding, scoring, recognition and chat requests/responses separately.
• Treat required repository code as a separate versioned dependency to review; do not enable remote code blindly to suppress errors.

The cards show different library examples. This guide does not invent an all-platform compatibility matrix. Combinations without installation checks must not be called tested-compatible.

• 维度/索引不匹配：先核对嵌入版本、模式与索引元数据，再看 工具参数排查。
• 召回找不到：检查语料、分块、过滤与查询，再看 检索排查。
• 图像/音频打不开：先看 文件排查。
• 内存不足：使用 上下文与并发预算，先确认公式适用范围。编码器、图像和音频管线不能无条件套解码器 KV 公式。

记录失败输入、环境版本、输出和核对位置，修改一个条件再复核；不要把模型卡中的演示结果写成自己的运行记录。

• Dimension/index mismatch: inspect embedding revision, mode and index metadata, then tool-input troubleshooting.
• Retrieval miss: inspect corpus, chunks, filters and query, then retrieval troubleshooting.
• Unreadable image/audio: start with file troubleshooting.
• Memory exhaustion: use the context/concurrency budget after confirming scope. Do not apply a decoder KV formula universally to encoders or image/audio pipelines.

Record input, environment versions, output and evidence; change one condition before rechecking. Never present a model-card demonstration as your own execution record.