QAnything

1. Project/Product Overview

2. What does it mostly do?

The core of the QAnything is a complete document parsing → vector indexing → two-stage retrieval → LLM generation pipeline. Users only need to upload files to obtain accurate answers based on document content.

2.1 Document Analysis Ability (Youdao Core Accumulation)

2.2 OCR Capability

The technical accumulation of Youdao Translation/OCR is the core barrier that distinguishes QAnything from other domestic RAG schemes. For Chinese scanned PDF (image PDF), the OCR recognition accuracy of the QAnything is significantly better than that of the general OCR engine. In v2.0, OCR can be called separately as a separate service (HTTP).

115.00g Phase Retrieval (Core Architecture Differentiation)

QAnything uses the self-developed BCEmbedding model (Apache-2.0 open source) and consists of two components:

- One-stage Embedding(bce-embedding-base_v1) :MTEB has a comprehensive score of 59.43, significantly better than BGE-large-zh-v1.5(54.21) and M3E-base(53.54), especially in Chinese cross-language scenes.

-Phase II Rerank(bce-reranker-base_v1):Reranking score 60.06, better than BGE-reranker-large(59.69)

-Combination effect SOTA: In the LlamaIndex RAG evaluation, the combination of BCEmbedding BCReranker reaches the current best.

The core value of two-stage retrieval: the larger the amount of data in the knowledge base, the problem of "retrieval degradation" will appear in one-stage Embedding retrieval, and Rerank rearrangement can reverse the trend and realize "the more data, the better the effect".

2.4 Hybrid Search & Other Features

3. Applicable Scenario

Report study/investment research, content summary, key point extraction, document question and answer, investment institutions, consulting companies.

4. Not quite the scene

5. Core Competence List

6. Architecture/deployment/integration approach

Overall Architecture

用户上传文档（PDF/Word/PPT/...）
    │
    ▼
文档解析层（PDF Parser / OCR / 格式转换器）
    │
    ▼
文本分块 + 元数据提取
    │
    ▼
向量化索引（Embedding 服务 + Elasticsearch/Milvus 等）
    │
    ▼
用户提问 → 一阶段 Embedding 检索（粗筛）
    │
    ▼
二阶段 Rerank 重排序（精排）
    │
    ▼
LLM 生成回答（OpenAI 兼容接口）
    │
    ▼
返回答案 + 溯源引用

Deployment Mode

Hardware Requirements

Model Integration

-LLM: all OpenAI API-compatible models (Ollama, Tongyi Qiwen DashScope, DeepSeek, GLM, etc.)

-Embedding Rerank: default BCEmbedding (self-developed, replaceable)

-Vector storage:Elasticsearch (built-in Chinese word segmentation IK)

API Support

Provides RESTful API, which can perform all operations such as file upload, knowledge base management, and Q & A. In v2.0, Embed, Rerank, OCR, and PDF parsing all support independent HTTP calls.

How to use #7.

Docker one-click deployment (recommended)

# 克隆项目
git clone https://github.com/netease-youdao/QAnything.git
cd QAnything

# 启动服务（根据操作系统选择 compose 文件）
# Linux
docker compose -f docker-compose-linux.yaml up -d

# Mac
docker compose -f docker-compose-mac.yaml up -d

# Windows
docker compose -f docker-compose-win.yaml up -d

# 访问 Web UI
# 浏览器打开 http://localhost:5052

Configure LLM

To set the page configuration in the Web UI frontend:

-'API_BASE': the API address of the LLM service (e. g. 'https://api.openai.com/v1' or Ollama's' http:// localhost:11434/v1')

-'API_KEY: API Key

-'MODEL': model name (e. g. gpt-4o, qwen-plus, deepseek-chat)

API call example

import requests

# 上传文件到知识库
url = "http://localhost:5052/api/local_doc_qa/upload_files"
files = {"files": open("contract.pdf", "rb")}
data = {"kb_id": "KB123456", "user_id": "user001"}
resp = requests.post(url, files=files, data=data)

# 问答
url = "http://localhost:5052/api/local_doc_qa/local_doc_chat"
payload = {
    "question": "这份合同的关键条款是什么？",
    "kb_ids": ["KB123456"],
    "user_id": "user001"
}
resp = requests.post(url, json=payload)
print(resp.json()["response"])

Use BCEmbedding independent components

# BCEmbedding 是 Apache-2.0 协议，可单独使用
from BCEmbedding import EmbeddingModel, RerankerModel

# Embedding
embed_model = EmbeddingModel(model_name_or_path="maidalun1020/bce-embedding-base_v1")
embeddings = embed_model.encode(["什么是RAG?", "RAG是检索增强生成"])

# Rerank
reranker = RerankerModel(model_name_or_path="maidalun1020/bce-reranker-base_v1")
scores = reranker.compute_score(["什么是RAG?"], ["RAG是检索增强生成技术"])

8. What can I say before sales

8.1 a sentence positioning

• * "QAnything is a document intelligent question and answer system produced by Netease Youdao-throw in the scanned Chinese document and give accurate answers in seconds. "**

8.2 customer pain points → solutions

8.3 Differentiated Selling Points

vs RAGFlow(InfiniFlow open source):

-QAnything advantages: Youdao OCR accumulation → better PDF analysis of Chinese scanned documents/forms; BCEmbedding bilingual and cross-language SOTA; Pure CPU deployment is lighter

-RAGFlow advantages: more comprehensive DeepDoc parsing engine (support for layout recognition YOLOv8); V0.21 introduces Ingestion Pipeline that can be arranged; The update is very active (2025-2026 continuous high frequency iteration);Apache-2.0 protocol is more friendly

-Conclusion: If the customer's core requirement is "a large number of complex documents in various formats can be Pipeline by deep analysis", select RAGFlow; If it is "simple deployment of Chinese scanned documents/forms PDF Q & A", the QAnything is more accurate.

vs MaxKB(1Panel open source):

-QAnything advantages: stronger retrieval model (BCEmbedding SOTA vs MaxKB basic retrieval), deeper document analysis (especially OCR/table), commercial support in enterprise version

-MaxKB Advantages: Stars More (20,600), Protocol Friendly (GPL-3.0 with MaxKB EULA), Workflow Orchestration and MCP Tool Call, Much Higher Community Activity, Deep Integration with 1Panel Operation and Maintenance Ecology

-Conclusion: MaxKB is more suitable as a general platform Agent base; QAnything focus more on the "document → question and answer" path to the extreme.

vs Dify:

-QAnything is "Documentation Q & A Special Tool";Dify is "AI Application Development Platform"

-Dify has visual workflow orchestration, rich tools and plug-in ecology, but the document analysis depth is not as deep as QAnything.

-QAnything is suitable for customers who do not need complicated choreography and only need document questions and answers; Dify is suitable for enterprises that need custom AI applications.

• Core Differentiation Summary: Youdao OCR BCEmbedding Pure CPU Deployment Scan Friendly *

8.4 Customer Value Story Line

1. Cut in *:"Your company has a lot of scanned contracts/files/reports, want to use AI Q & A but OCR effect is not good?"
2. Resonance :"The universal OCR engine has a low recognition rate for Chinese scanned documents, the contents of the form are broken after page spanning, and the column layout is misread-these are common pits."
3. Demo: Upload a Chinese scanned PDF contract (including forms), and ask "What is the liability clause for breach of contract?" -- QAnything accurately identify the scanned text, parse the forms, locate the answers, and trace the original text.
4. Advanced : Build Enterprise Knowledge Base from Single Document → Batch Upload → Custom Bot (Sales Assistant/Legal Assistant/Technical Consultant) → API Integration into OA System
5. Ends :"Netease has 20 years of OCR technology accumulation, 14,000 GitHub Stars,3000 WeChat community users-there is no stronger open source scheme on this subdivision track."

9. Frequently Asked Customer Questions

10. PoC Recommendations

Recommended PoC Direction: Chinese Scanned PDF Q & A

This is the core differentiation scenario of QAnything, and it is recommended that PoC focus on this to maximize its irreplaceability.

Validation Metrics:

-OCR text recognition accuracy> 95% (Chinese scan)

-Complete extraction rate of table data> 90%

-End-to-end answer accuracy> 85% (based on document factual validation)

-Average response time <5 seconds (CPU-only environment)

-Answer traceability is accurate (the location of the cited document matches the answer)

PoC Note:

-Be sure to use the customer's own real documents, not clean typeset test documents.

-If the customer has concerns about the AGPL protocol, the PoC phase should be clear: the open source version is only used for technical verification, and the commercial use requires the enterprise version authorization.

-Inform customers in advance of the expected indexing speed of large-scale documents in CPU-only mode

11. Risks and Considerations

12. My Pre-Sales Judgment

Recommendation: Cautiously recommended (very suitable for specific scenarios, but the overall risk is high)

Reason:

1. irreplaceable differentiation advantage : There is a combination of OCR BCEmbedding. At present, there is no better open source scheme for processing the subdivision scenario of Chinese scanned PDF. If the customer happens to be in this scenario, QAnything is the most preferred.
2. Deployment Friendly :Docker one-click startup, pure CPU operation, mirror image 4.88GB, extremely friendly to small and medium-sized enterprises without GPU.
3. Full-featured and stable: Although it is no longer updated, v2.0 is already a mature and fully functional document answering system.
4. But-AGPL and maintenance stagnation are two thunder : Legal Risk Technology Stagnation = Customers must be fully informed before sales and cannot be avoided.

Recommended Customer Persona:

-Core Appeal Accurately Hits "Chinese Scanned Document/Form PDF → Local Q & A"

-Pure intranet deployment, do not provide external document Q & A service (to avoid AGPL risks)

-The requirement for function update frequency is not high, and the existing functions of v2.0 can already meet the demand

-Limited budget but need OCR advantage (open source version can meet)

-Or have a budget to buy the enterprise version (more powerful business license)

Not recommended situations:

-Customer Legal Affairs explicitly prohibits the use of AGPL protocol → Directly exclude open source version, push enterprise version or change RAGFlow/MaxKB

-Need long-term continuous function iteration and community support → recommend RAGFlow (update the most active)

-Workflow orchestration, multi-Agent, plug-in ecology required → Dify/MaxKB recommended

-Documents are mainly in English or non-scanned documents → OCR advantage is not obvious, RAGFlow or Haystack is recommended

-High concurrency external services (and do not want to pay) → AGPL risk unacceptable

Pre-sales strategy recommendations:

-First judge the type of customer document: there are a large number of scanned documents → QAnything is the first recommendation

-reconfirm AGPL's position: legal OK and intranet use → direct push of open source PoC

-Legal service is not OK or needs external service → Push enterprise version or BCEmbedding self-built scheme (BCEmbedding Apache-2.0!)

13. REFERENCE

-GitHub repository: https://github.com/netease-youdao/QAnything

-official website/online experience: https://qanything.ai

-BCEmbedding (Retrieval Model, Apache-2.0):https://github.com/netease-youdao/BCEmbedding

-Youdao Speed Reading (online trial):https://read.youdao.com

-FAQ (Chinese):https://github.com/netease-youdao/QAnything/blob/qanything-v2/FAQ_zh.md

-Demand feedback: https://qanything.canny.io/feature-requests

-HuggingFace model: https://huggingface.co/maidalun1020

-Enterprise Business Contact: AIcloud_Business@corp.youdao.com -8255-8901

• Analysis Date: 2026-07-02 | Data Aging: GitHub Information Pull in Real Time, Official Website Content from qanything. AI, Competition Comparison Based on 2026 Latest Data *