Ask or search…

Technical Spec

For those already familiar with LLM application tech stacks, this document serves as a shortcut to understand Dify's unique advantages
We adopt transparent policies around product specifications to ensure decisions are made based on complete understanding. Such transparency not only benefits your technical selection, but also promotes deeper comprehension within the community for active contributions.

Project Basics

March 2023
Open Source License
Official R&D Team
Over 10 full-time employees
Community Contributors
Over 70 people
Backend Technology
Frontend Technology
Codebase Size
Over 130,000 lines
Release Frequency
Average once per week

Technical Features

LLM Inference Engines
Dify Runtime (LangChain removed since v0.4)
Commercial Models Supported
10+, including OpenAI and Anthropic Onboard new mainstream models within 48 hours
MaaS Vendor Supported
2, Hugging Face and Replicate
Local Model Inference Runtimes Supported
5, Xoribits (recommended), OpenLLM, LocalAI, ChatGLM,Ollama
Multimodal Capabilities
ASR Models
Rich-text models up to GPT-4V specs
Built-in App Types
Text generation, Conversational
Prompt-as-a-Service Orchestration
Visual orchestration interface widely praised, modify Prompts and preview effects in one place.
Orchestration Modes
  • Simple orchestration
  • Advanced orchestration
  • Assistant orchestration
  • Flow orchestration (Q1 2024)
Prompt Variable Types
  • String
  • Radio enum
  • External API
  • File (Jan 2024)
RAG Features
Industry-first visual knowledge base management interface, supporting snippet previews and recall testing.
Indexing Methods
  • Keywords
  • Text vectors
  • LLM-assisted question-snippet model
Retrieval Methods
  • Keywords
  • Text similarity matching
  • N choose 1
  • Multi-path recall
Recall Optimization
  • Re-rank models
ETL Capabilities
Automated cleaning for TXT, Markdown, PDF, HTML, DOC, CSV formats. Unstructured service enables maximum support.
Sync Notion docs as knowledge bases.
Vector Databases Supported
Qdrant (recommended), Weaviate, Zilliz
Agent Technologies
ReAct, Function Call.
Tooling Support
  • Invoke OpenAI Plugin standard tools
  • Directly load OpenAPI Specification APIs as tools
Built-in Tools
  • 11 tools
Supported, annotations based on logs
Annotation Reply
Based on human-annotated Q&As, used for similarity-based replies. Exportable as data format for model fine-tuning.
Content Moderation
OpenAI Moderation or external APIs
Team Collaboration
Workspaces, multi-member management
API Specs
RESTful, most features covered
Deployment Methods
Docker, Helm