Diting AI
Listen to Everything · Answer the Future
An enterprise-grade intelligent knowledge platform built on RAG and AI Agents.
Hybrid retrieval, four-level evidence evaluation, citation tracing — every answer is verifiable.
Open Source ▾
According to the 2024 Annual Business Report, East China generated revenue of ¥420M, accounting for 38.6% of total revenue, up 12.3% YoY…▌
Product Editions
One knowledge engine, two delivery forms
Diting Web
Web-based RAG platform — zero deployment, instant access, with online knowledge base management and smart chat.
- ✓Ready to use, no installation
- ✓Multi-device sync, access anywhere
- ✓Team collaboration with permissions
- ✓Automatic cloud updates
Diting Desktop
Desktop client with local execution for data security, offline usage and local hardware acceleration.
- ✓Runs locally, data never leaves
- ✓Offline capable, no network required
- ✓GPU acceleration, faster response
- ✓System integration, global shortcuts
Six Core Innovations
In-house engine capabilities beyond off-the-shelf RAG solutions
LLM Query Planning Engine
The LLM analyzes intent and dynamically selects DIRECT retrieval, REWRITE optimization, or DECOMPOSE parallel strategies — complex questions are split into sub-queries, retrieved in parallel, then aggregated.
Three-Layer Hybrid Retrieval
PGVector semantic search + ES BM25 keyword search, fused and ranked by RRF, then refined through cluster aggregation and neighbor-window expansion to complete cross-chunk context.
Four-Level Evidence Evaluation
Pre-generation gating: NONE / WEAK triggers a polite refusal, PARTIAL answers with caveats, SUFFICIENT generates normally with citation tracing — no more confident hallucinations.
ReactAgent Autonomy
A graph execution engine drives "Think → Tool Call → Respond". CHAT / KB_SEARCH modes switch dynamically within one session, with SSE streaming output.
Three-Level Memory Compression
Session memory → compact summary → runtime truncation: progressive compression that maximizes long-dialogue coherence within a limited context window.
Enterprise Security & Isolation
RBAC permissions + group-level data isolation, RSA/AES dynamic API encryption, field-level encryption (AES/RSA/SM2/SM4), and full operation auditing.
End-to-End RAG Workflow
A complete pipeline from document upload to citation tracing
Upload
Chunked upload · resumable · instant dedup · SHA-256 verification
→ETL Pipeline
Tika parsing → cleansing → structure-aware chunking (500/800/80 tokens)
→Query Planning
LLM intent analysis, auto routing via DIRECT / REWRITE / DECOMPOSE
→Hybrid Retrieval
Semantic + keyword channels → RRF fusion → cluster aggregation → neighbor expansion
→Evidence Evaluation
Four-level sufficiency gating; refuses to answer when evidence is insufficient
→Generation & Tracing
SSE streaming output with cited chunks, source documents and relevance scores
Tech Stack
Enterprise-grade Java full stack — mature and stable choices
Why Diting
Compared with LangChain / LlamaIndex / Dify / FastGPT
Listen to Everything, Answer the Future
Combine an enterprise admin framework with RAG technology to build your own intelligent knowledge platform.