bailian-docs-llm-wiki

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

百炼文档知识库

Tongyi Qianfan Document Knowledge Base

阿里云百炼平台的完整技术文档知识库，涵盖模型使用、应用开发、API 参考等内容。

The complete technical document knowledge base for Alibaba Cloud Tongyi Qianfan Platform, covering model usage, application development, API references, and other content.

何时使用

When to Use

当用户涉及以下场景时激活此 Skill：

查询百炼平台的模型列表、模型参数、调用方式（含具体模型的 contextWindow、QPM、定价、sample code 等结构化字段 → 查
```
models/
```
）
查阅百炼 API 参数、请求/响应格式、错误码
了解百炼应用开发（智能体、RAG、知识库、记忆、插件等）
选择模型、比较模型能力、了解模型定价与限流
使用百炼 SDK / OpenAI 兼容接口
语音识别、语音合成、图片生成、视频生成等多模态能力
Token Plan、计费、免费额度等商务问题

Activate this Skill when users are involved in the following scenarios:

Querying Qianfan platform's model lists, model parameters, calling methods (including structured fields like contextWindow, QPM, pricing, sample code for specific models → check
```
models/
```
)
Looking up Qianfan API parameters, request/response formats, error codes
Learning about Qianfan application development (Agent, RAG, Knowledge Base, Memory, Plugins, etc.)
Selecting models, comparing model capabilities, understanding model pricing and rate limits
Using Qianfan SDK / OpenAI compatible interfaces
Multimodal capabilities such as speech recognition, speech synthesis, image generation, video generation
Business issues like Token Plan, billing, free quota, etc.

何时不要使用

When Not to Use

下面这些场景与百炼文档无关，不要激活本 Skill：

询问 OpenAI / Anthropic / Google 等其他厂商的 API、模型或定价
通用编程问题、框架问题（React、Vue、Spring 等）
```
bl
```
CLI 命令本身的用法（那是
```
bailian-cli
```
skill 的职责）
与百炼无关的阿里云产品（OSS、ECS、RDS 等）

Do not activate this Skill for the following scenarios that are unrelated to Qianfan documents:

Inquiring about APIs, models or pricing of other vendors like OpenAI / Anthropic / Google
General programming issues, framework issues (React, Vue, Spring, etc.)
Usage of the
```
bl
```
CLI command itself (that's the responsibility of the
```
bailian-cli
```
skill)
Alibaba Cloud products unrelated to Qianfan (OSS, ECS, RDS, etc.)

文档层级

Document Layers

1. Wiki 层（合成层） —

wiki/

1. Wiki Layer (Synthesis Layer) —

wiki/

由 LLM 自动合成的结构化 Wiki，包含三类页面：

主题页：按功能领域聚合的综合文档（
```
wiki/guides/*.md
```
、
```
wiki/api/*.md
```
）
概念页：跨主题的横切概念（
```
wiki/concepts/*.md
```
，如 RAG、流式输出、Function Calling 等）
对比页：同类方案的结构化对比分析（
```
wiki/comparisons/*.md
```
，含对比表格）

优先查阅
wiki/index.md
，它是全部 wiki 页面的索引入口。

每篇主题页末尾都有

## 来源文档

段落，列出对应的 raw 原文路径，可用于追溯和深度阅读。

Structured Wiki automatically synthesized by LLM, including three types of pages:

Topic Pages: Comprehensive documents aggregated by functional domains (
```
wiki/guides/*.md
```
,
```
wiki/api/*.md
```
)
Concept Pages: Cross-cutting concepts across topics (
```
wiki/concepts/*.md
```
, such as RAG, streaming output, Function Calling, etc.)
Comparison Pages: Structured comparative analysis of similar solutions (
```
wiki/comparisons/*.md
```
, including comparison tables)

Prioritize checking
wiki/index.md
, which is the index entry for all wiki pages.

Each topic page ends with a

## Source Documents

section, listing the corresponding raw original document paths for traceability and in-depth reading.

2. Models 层（模型市场结构化数据） —

models/

2. Models Layer (Structured Data of Model Market) —

models/

直接从百炼 console 网关

listFoundationModels

抓取的结构化模型元数据，对应「模型广场」页面。涉及具体模型的能力、上下文长度、QPM 限流、定价、官方 sample code 等问题时，优先查这里，比 wiki 更新更准。

三层产物，单一数据源（

families.jsonl

models.jsonl

是机器查询入口，

index.md

是

families.jsonl

的人类可读视图）：

```
models/index.md
```
— 按主能力分桶的家族索引（中文标签 + 模型代码 + 链接到明细 JSON），适合「概览扫一眼」

models/families.jsonl

— 每行一个家族，含

slug

name

description

primaryCapability

capabilities

（家族下所有 item 的并集）/

providers

itemCount

maxContextWindow

/ 轻量

items[]

摘要（每个 item 只含

model

name

contextWindow

capabilities

openSource

）/

detailPath

。适合按家族筛选，例如：

列出推理类家族：

jq -c 'select(.primaryCapability=="Reasoning") | {slug,name,itemCount}' families.jsonl

找包含某个 provider 的家族：

grep '"moonshot-ai"' families.jsonl | jq -c .slug

models/models.jsonl

— 每行一个主干模型（跨家族扁平），含

model

family

capabilities

features

contextWindow

maxInputTokens

maxOutputTokens

prices

（精简）/

qpmInfo

（精简）/

docUrl

detailPath

。适合跨家族批量查询，例如：

列出所有 contextWindow ≥ 1M 的模型：

jq -c 'select(.contextWindow>=1000000)' models.jsonl

找支持 function-calling 的文本模型：

grep '"function-calling"' models.jsonl | jq -c '{model,family,contextWindow}'

两份 JSONL 的 join 字段：

models.jsonl[].family == families.jsonl[].slug

。命中后按

detailPath

打开

groups/<slug>.json

取完整字段（

samples

、

predictConfig

）。

```
models/groups/<slug>.json
```
— 单个模型家族（如
```
qwen3-max
```
、
```
deepseek
```
、
```
wan-image-to-video
```
）的完整明细：
- 家族层：
```
name
```
  、
```
description
```
- ```
items[]
```
  ：该家族下所有主干模型版本（剔除了带日期后缀的快照），已经过裁剪（丢弃了
```
permissions
```
  /
```
activationStatus
```
  /
```
quota
```
  /
```
scope
```
  /
```
tags
```
  /
```
license
```
  /
```
serviceSites
```
  /
```
modelInfo
```
  等与模型能力无关的账号/UI 字段），保留：
```
model
```
  （API 调用名）、
```
modelAlias
```
  、
```
contextWindow
```
  、
```
maxInputTokens
```
  、
```
maxOutputTokens
```
  、
```
capabilities
```
  、
```
features
```
  、
```
provider
```
  、
```
docUrl
```
  、
```
prices
```
  、
```
qpmInfo
```
  、
```
samples
```
  （调用示例，扁平化结构：
```
samples.<sdk>.<api>.{curl,python,nodejs,java,docUrl}
```
  ，例如
```
samples.openai.completionsAPI.python
```
  直接是代码字符串）、
```
predictConfig
```
  （模型调用入参定义：从
```
getPredictParamConfig
```
  网关获取，含
```
system
```
  /
```
temperature
```
  /
```
top_p
```
  /
```
enable_search
```
  /
```
enable_thinking
```
  等参数的
```
name
```
  /
```
key
```
  /
```
default
```
  /
```
tip
```
  /
```
range
```
  ，与百炼 Playground 对齐）
```
models/meta.json
```
— 增量爬取的指纹缓存，不要给用户引用

数据通过

pnpm --filter bailian-docs-llm-wiki run crawl:models

刷新。

Structured model metadata directly crawled from the Qianfan console gateway

listFoundationModels

, corresponding to the "Model Plaza" page. For issues related to specific model capabilities, context length, QPM rate limits, pricing, official sample code, etc., check here first as it is more up-to-date and accurate than the wiki.

Three-layer products, single data source (

families.jsonl

models.jsonl

are machine query entries,

index.md

is a human-readable view of

families.jsonl

```
models/index.md
```
— Family index grouped by primary capabilities (Chinese labels + model codes + links to detailed JSON), suitable for "quick overview"

models/families.jsonl

— One family per line, including

slug

name

description

primaryCapability

capabilities

(union of all items under the family) /

providers

itemCount

maxContextWindow

/ lightweight

items[]

summary (each item only contains

model

name

contextWindow

capabilities

openSource

) /

detailPath

. Suitable for filtering by family, for example:

List reasoning families:

jq -c 'select(.primaryCapability=="Reasoning") | {slug,name,itemCount}' families.jsonl

Find families containing a specific provider:

grep '"moonshot-ai"' families.jsonl | jq -c .slug

models/models.jsonl

— One backbone model per line (flat across families), including

model

family

capabilities

features

contextWindow

maxInputTokens

maxOutputTokens

prices

(simplified) /

qpmInfo

(simplified) /

docUrl

detailPath

. Suitable for cross-family batch queries, for example:

List all models with contextWindow ≥ 1M:

jq -c 'select(.contextWindow>=1000000)' models.jsonl

Find text models that support function-calling:

grep '"function-calling"' models.jsonl | jq -c '{model,family,contextWindow}'

Join field for the two JSONL files:

models.jsonl[].family == families.jsonl[].slug

. After matching, open

groups/<slug>.json

via

detailPath

to get complete fields (

samples

predictConfig

```
models/groups/<slug>.json
```
— Complete details of a single model family (such as
```
qwen3-max
```
,
```
deepseek
```
,
```
wan-image-to-video
```
):
- Family layer:
```
name
```
  ,
```
description
```
- ```
items[]
```
  : All backbone model versions under the family (excluding snapshots with date suffixes), which have been trimmed (discarded account/UI fields irrelevant to model capabilities such as
```
permissions
```
  /
```
activationStatus
```
  /
```
quota
```
  /
```
scope
```
  /
```
tags
```
  /
```
license
```
  /
```
serviceSites
```
  /
```
modelInfo
```
  ), retaining:
```
model
```
  (API call name),
```
modelAlias
```
  ,
```
contextWindow
```
  ,
```
maxInputTokens
```
  ,
```
maxOutputTokens
```
  ,
```
capabilities
```
  ,
```
features
```
  ,
```
provider
```
  ,
```
docUrl
```
  ,
```
prices
```
  ,
```
qpmInfo
```
  ,
```
samples
```
  (call examples, flattened structure:
```
samples.<sdk>.<api>.{curl,python,nodejs,java,docUrl}
```
  , e.g.,
```
samples.openai.completionsAPI.python
```
  is directly a code string),
```
predictConfig
```
  (model call parameter definition: obtained from the
```
getPredictParamConfig
```
  gateway, including
```
name
```
  /
```
key
```
  /
```
default
```
  /
```
tip
```
  /
```
range
```
  for parameters like
```
system
```
  /
```
temperature
```
  /
```
top_p
```
  /
```
enable_search
```
  /
```
enable_thinking
```
  , aligned with Qianfan Playground)
```
models/meta.json
```
— Fingerprint cache for incremental crawling, do not reference this for users

Data is refreshed via

pnpm --filter bailian-docs-llm-wiki run crawl:models

能力代码（capability）映射

Capability Code Mapping

index.md

的章节标题与

items[].capabilities[]

用的是英文短代码，对应中文含义如下：

代码	中文标签
`TG`	文本生成
`Reasoning`	推理
`VU`	视觉理解
`IG`	图像生成
`VG`	视频生成
`TTS`	语音合成
`ASR`	语音识别
`Realtime-ASR`	实时语音识别
`Realtime-Text-to-Speech`	实时语音合成
`Realtime-Audio-Translate`	实时音频翻译
`Realtime-Omni`	实时全模态
`Multimodal-Omni`	全模态
`ME`	多模态嵌入
`TR`	翻译
`3D-generation`	3D 生成

一个模型常常带多个 capability，

index.md

中按

capabilities[0]

（主能力）归类，查找时按中文标签即可定位章节。

The section titles in

index.md

and

items[].capabilities[]

use English short codes, corresponding to Chinese meanings as follows:

Code	Chinese Label
`TG`	Text Generation
`Reasoning`	Reasoning
`VU`	Visual Understanding
`IG`	Image Generation
`VG`	Video Generation
`TTS`	Speech Synthesis
`ASR`	Speech Recognition
`Realtime-ASR`	Realtime Speech Recognition
`Realtime-Text-to-Speech`	Realtime Text-to-Speech
`Realtime-Audio-Translate`	Realtime Audio Translation
`Realtime-Omni`	Realtime Multimodal
`Multimodal-Omni`	Multimodal
`ME`	Multimodal Embedding
`TR`	Translation
`3D-generation`	3D Generation

A model often has multiple capabilities. In

index.md

, models are categorized by

capabilities[0]

(primary capability). You can locate the section by Chinese label during search.

3. Raw 层（原始层） —

raw/

3. Raw Layer (Original Layer) —

raw/

从 help.aliyun.com 爬取的原始文档，按分类存放：

```
raw/model-user-guide/
```
— 模型使用指南
```
raw/application-user-guide/
```
— 应用使用指南
```
raw/model-api-reference/
```
— 模型 API 参考
```
raw/application-api-reference/
```
— 应用 API 参考

当 wiki / models 层信息不足时，从 raw 层获取完整原文。

Original documents crawled from help.aliyun.com, stored by category:

```
raw/model-user-guide/
```
— Model User Guide
```
raw/application-user-guide/
```
— Application User Guide
```
raw/model-api-reference/
```
— Model API Reference
```
raw/application-api-reference/
```
— Application API Reference

When information in the wiki / models layers is insufficient, obtain complete original text from the raw layer.

强约束

Strong Constraints

回答必须基于
wiki/
或
raw/
下的实际文件内容，不得凭印象编造 API、参数名、错误码或定价数字。
引用时附带相对路径（如
```
wiki/api/qwen-api.md
```
），方便用户验证。
如果在
```
wiki-metadata.json
```
中查到某页面
```
qualityScore <= 2
```
，直接绕过 wiki 该页，回退到
```
raw/
```
原文回答。

Answers must be based on actual file content under
wiki/
or
raw/
, do not fabricate APIs, parameter names, error codes or pricing figures from memory.
Attach relative paths when referencing (e.g.,
```
wiki/api/qwen-api.md
```
) to facilitate user verification.
If a page has
```
qualityScore <= 2
```
found in
```
wiki-metadata.json
```
, directly bypass this wiki page and fall back to the original text in
```
raw/
```
for answering.

查阅流程

Lookup Process

模型规格类问题先查 models：
- 具体某个模型的能力、上下文、QPM、价格、官方示例代码 → 看
```
index.md
```
  定位家族 → 打开
```
groups/<slug>.json
```
- 跨家族筛选模型（如「上下文 ≥ 1M 的模型」「支持 function-calling 的推理模型」）→ grep
```
models.jsonl
```
- 按家族维度筛选（如「有哪些纯推理家族」「moonshot 旗下都有哪些家族」）→ grep
```
families.jsonl
```
- 命中后按
```
detailPath
```
  打开对应
```
groups/<slug>.json
```
  取
```
samples
```
  /
```
predictConfig
```
  等完整字段
概念/使用方式查 wiki 索引：读取
```
wiki/index.md
```
找到对应的主题页、概念页或对比页
概念/对比优先：跨领域问题或方案选型问题，优先看
```
wiki/concepts/*
```
与
```
wiki/comparisons/*
```
回溯 raw 原文：从主题页末尾的
```
## 来源文档
```
列表点进
```
raw/.../*.md
```
看完整细节
全文索引：
```
llms.txt
```
含完整目录树；
```
llms-full.txt
```
含全文拼接，可用 grep
特定参数/错误码：直接在
```
raw/
```
下用 grep

Check models first for model specification issues:
- Capabilities, context, QPM, pricing, official sample code of a specific model → Locate the family via
```
index.md
```
  → Open
```
groups/<slug>.json
```
- Cross-family model filtering (e.g., "models with context ≥ 1M", "reasoning models that support function-calling") → grep
```
models.jsonl
```
- Family-level filtering (e.g., "which pure reasoning families are there", "what families are under moonshot") → grep
```
families.jsonl
```
- After matching, open the corresponding
```
groups/<slug>.json
```
  via
```
detailPath
```
  to get complete fields like
```
samples
```
  /
```
predictConfig
```
Check wiki index for concepts/usage methods: Read
```
wiki/index.md
```
to find the corresponding topic pages, concept pages or comparison pages
Prioritize concepts/comparisons: For cross-domain issues or solution selection issues, prioritize checking
```
wiki/concepts/*
```
and
```
wiki/comparisons/*
```
Trace back to raw original text: Access
```
raw/.../*.md
```
from the
```
## Source Documents
```
list at the end of topic pages to view complete details
Full-text index:
```
llms.txt
```
contains the complete directory tree;
```
llms-full.txt
```
contains full-text concatenation, which can be searched with grep
Specific parameters/error codes: Directly use grep under
```
raw/
```

快速定位映射

Quick Location Mapping

关键词	推荐入口路径
某模型的 contextWindow / 价格 / QPM / sample code / 入参定义	`models/groups/<slug>.json` （从 `models/index.md` 找 slug）
跨家族筛选模型：按 contextWindow / capability / feature / price 批量查找	`models/models.jsonl` （ `grep` / `jq` 一行一模型）
按家族筛选：按 primaryCapability / providers / itemCount / maxContextWindow 找家族	`models/families.jsonl` （一行一家族，含 items[] 摘要）
模型家族总览 / 按能力分桶浏览	`models/index.md`
模型列表 / Qwen / DeepSeek	`wiki/guides/more-about-models.md`
文本对话 / Chat Completion	`wiki/api/text-generation-api-reference.md`
语音合成 / TTS	`wiki/api/speech-synthesis-api-reference.md`
语音识别 / ASR	`wiki/api/speech-recognition-api-reference.md`
图片生成 / 图片编辑	`wiki/api/image-generation-api-reference.md`
视频生成	`wiki/api/video-generation-api-reference.md`
多模态 / Omni	`wiki/api/multimodal-api-reference.md`
Function Calling / 工具调用	`wiki/concepts/function-calling.md`
RAG / 知识库	`wiki/concepts/rag.md` + `wiki/guides/knowledge-base.md`
智能体 / 应用调用	`wiki/guides/application-call.md`
OpenAI 兼容	`wiki/concepts/openai-compatibility.md`
Token / 计费 / 限流	`wiki/guides/billing-and-rate-limit.md`

实际文件名以
wiki/index.md
为准；上表若有出入应回到索引页查找。

Keywords	Recommended Entry Path
contextWindow / pricing / QPM / sample code / parameter definition of a model	`models/groups/<slug>.json` (find slug from `models/index.md` )
Cross-family model filtering: Batch search by contextWindow / capability / feature / price	`models/models.jsonl` (one model per line, use `grep` / `jq` )
Family-level filtering: Find families by primaryCapability / providers / itemCount / maxContextWindow	`models/families.jsonl` (one family per line, includes items[] summary)
Overview of model families / browsing by capability buckets	`models/index.md`
Model list / Qwen / DeepSeek	`wiki/guides/more-about-models.md`
Text chat / Chat Completion	`wiki/api/text-generation-api-reference.md`
Speech synthesis / TTS	`wiki/api/speech-synthesis-api-reference.md`
Speech recognition / ASR	`wiki/api/speech-recognition-api-reference.md`
Image generation / image editing	`wiki/api/image-generation-api-reference.md`
Video generation	`wiki/api/video-generation-api-reference.md`
Multimodal / Omni	`wiki/api/multimodal-api-reference.md`
Function Calling / Tool Calling	`wiki/concepts/function-calling.md`
RAG / Knowledge Base	`wiki/concepts/rag.md` + `wiki/guides/knowledge-base.md`
Agent / Application Calling	`wiki/guides/application-call.md`
OpenAI Compatibility	`wiki/concepts/openai-compatibility.md`
Token / Billing / Rate Limit	`wiki/guides/billing-and-rate-limit.md`

Actual file names are subject to
wiki/index.md
; if there is any discrepancy in the above table, return to the index page to search.

注意事项

Notes

文档持续更新，如遇矛盾以
```
models/
```
>
```
raw/
```
>
```
wiki/
```
的顺序为准（
```
models/
```
直接拉自 console 网关，最新；
```
raw/
```
是源站爬取；
```
wiki/
```
是 LLM 合成可能滞后）
```
metadata.json
```
记录每篇 raw 文档的最后修改时间
```
wiki-metadata.json
```
记录每个 wiki 页面的合成 hash、时间戳、可选的
```
qualityScore
```
```
wiki/eval-report.md
```
是最近一次评测报告（结构问题、断链、长度异常等）

wiki/

目录可能尚未生成（需运行

pnpm --filter bailian-docs-llm-wiki run synthesize

）

Documents are continuously updated. In case of conflicts, follow the priority order:
```
models/
```
>
```
raw/
```
>
```
wiki/
```
(
```
models/
```
is directly pulled from the console gateway, the most up-to-date;
```
raw/
```
is crawled from the source site;
```
wiki/
```
is synthesized by LLM and may be lagging)
```
metadata.json
```
records the last modification time of each raw document
```
wiki-metadata.json
```
records the synthesis hash, timestamp, and optional
```
qualityScore
```
of each wiki page
```
wiki/eval-report.md
```
is the latest evaluation report (structural issues, broken links, length anomalies, etc.)
The
```
wiki/
```
directory may not have been generated yet (need to run
```
pnpm --filter bailian-docs-llm-wiki run synthesize
```
)