Introduction
"Bi Sheng invented movable type printing. BISHENG helps enterprises 'typeset' large language model capabilities into their own business processes with the same kind of flexible, recomposable modularity."
This is article #128 in the Open Source Project of the Day series. Today's project is BISHENG — DataElem's open-source enterprise LLM application DevOps platform.
Enterprises adopting LLMs frequently encounter a painful fragmentation: one tool specializes in RAG, another in workflow orchestration, another in model management, and yet another in permissions. Each solves one piece. Stitching them together creates a maintenance burden that compounds over time.
BISHENG positions as a one-stop solution: document parsing, knowledge base construction, RAG retrieval, visual workflow orchestration, multi-agent collaboration, model fine-tuning management, and enterprise permissions governance — all in a single platform.
What You'll Learn
- AGL (Agent Guidance Language) framework: how to encode domain expert experience into AI agents
- Visual workflow builder's Human-in-the-Loop design
- High-precision document parsing: OCR trained on 5+ years of proprietary data
- Enterprise RAG use cases: from policy comparison to resume screening
- BISHENG's technical architecture: Milvus + Elasticsearch + OnlyOffice + LLaMA-Factory
- v2.6.0's Linsight Task Mode
Prerequisites
- Familiarity with enterprise LLM application concepts (RAG, Agent, workflow)
- Experience with or interest in enterprise AI system development
- Docker basics (for local deployment)
Project Background
Why "Bi Sheng"
The project is named after Bi Sheng, the 11th-century inventor of movable type printing. The metaphor is fitting: where woodblock printing was rigid and inflexible, movable type was modular and recomposable. BISHENG aims to bring the same flexibility to enterprise AI capability building — not hardcoded fixed pipelines, but modular, orchestratable, iterable components.
Author / Team
- Company: DataElem (Shanghai)
- License: Apache-2.0
- Version: v2.6.0 (July 2026)
- User base: Fortune 500 companies and industry-leading organizations
Project Stats
- ⭐ GitHub Stars: 11,600+
- 🍴 Forks: 1,900+
- 📦 Releases: 73
- 📄 License: Apache-2.0
- 💻 Language: Python ~50% + TypeScript ~47.5%
Core Features
Linsight Agent (AGL Framework)
AGL (Agent Guidance Language) is BISHENG's proprietary agent guidance framework, addressing a core enterprise AI problem: how do you encode a specific domain expert's experience, preferences, and business logic into an AI, so it operates at expert level in professional scenarios?
A generic LLM performing financial analysis might produce technically correct but business-noncompliant conclusions. AGL lets you explicitly embed compliance specialists' judgment logic, common business edge cases, and industry-specific analytical perspectives into the agent's decision process.
Linsight Task Mode (v2.6.0): Agents don't just answer questions — they can proactively plan and execute multi-step tasks, delivering formatted reports that meet enterprise requirements.
Visual Workflow Builder
This is BISHENG's most differentiating core capability:
Single framework for all task types. Many platforms separate "chatflow" and "workflow" into independent modules with incompatible interfaces. BISHENG uses one visual workflow engine for every scenario.
Supported flow control:
- Loop: Agent repeats until a condition is met
- Parallelism: Multiple branches execute simultaneously
- Batch processing: Execute the same flow for each item in a list
- Conditional branching: Route to different paths based on intermediate results
Human-in-the-Loop:
An essential capability for enterprise scenarios:
Workflow running...
↓
Reaches "Contract Risk Assessment" node
↓
Legal Agent outputs preliminary assessment
↓
[PAUSE — waiting for human review]
↓ Human reviews, approves / modifies / rejects
↓
Continue with subsequent steps
This design inserts human judgment at high-risk decision points, rather than running fully automated from end to end. For enterprises with strict compliance requirements (finance, legal, healthcare), this is a hard requirement, not a nice-to-have.
High-Precision Document Parsing
BISHENG has a distinctive advantage in document parsing — not an integrated third-party tool, but a proprietary model DataElem trained on 5+ years of proprietary data:
Coverage:
- Printed text recognition (common and rare characters)
- Handwritten text recognition (including Chinese handwriting)
- Complex table recognition (including merged cells across rows and columns)
- Document layout analysis (multi-column, embedded images, header/footer separation)
- Seal/stamp recognition
- Chart and figure recognition
Deployment: Fully private deployment, no external API calls required, data never leaves the intranet.
This capability is particularly valuable in Chinese enterprise document processing: vast amounts of scanned paper forms, handwritten approval slips, and stamped documents are persistent pain points for traditional OCR tools. BISHENG's models have domain-specific optimization for exactly this type of data.
Enterprise RAG Knowledge Base
Complete enterprise knowledge base pipeline:
Source documents (PDF/Word/Excel/scans)
↓
Document parsing (high-precision OCR + layout analysis)
↓
Text chunking + vectorization (Milvus)
↓
Hybrid retrieval (vector similarity + full-text search via Elasticsearch)
↓
Knowledge base Q&A / citation tracing
Typical enterprise scenarios:
| Scenario | Description |
|---|---|
| Policy comparison | Clause-by-clause diff of new vs. old policy versions, generating change reports |
| Contract review | Automated comparison of contract terms against company standards, flagging risk points |
| Meeting minutes | Audio/video → structured minutes, automatic action item tracking |
| Resume screening | Multi-dimensional matching of candidate resumes against job requirements, scoring |
| Call record analysis | Structured analysis of customer service calls, key information extraction |
| Unstructured data governance | Historical document classification, tagging, knowledge graph construction |
Unified Model Management
Full lifecycle management from model selection to production operation:
- Model integration: OpenAI-compatible API / local deployment (llama.cpp, vLLM, etc.)
- Dataset management: Version control and quality control for labeled data
- SFT fine-tuning: LLaMA-Factory integration for fine-tuning on private data
- Evaluation: Build evaluation sets, quantify model performance on business scenarios
- Deployment management: Different departments using different model versions, traffic control, A/B testing
Enterprise-Grade Features
Permissions and security:
- RBAC (role-based access control)
- User group management (different permissions per department)
- SSO single sign-on (LDAP support)
- Per-group traffic control quotas
Operations and compliance:
- High-availability deployment
- Vulnerability scanning and security patching
- Usage statistics and audit logs
- Observability monitoring
Document editing:
- OnlyOffice integration for online collaborative editing of Word/Excel/PPT
- Fits the "AI generates draft → human edits online → final output" workflow pattern
Quick Deployment
System requirements:
- CPU: ≥ 4 cores (18 vCPUs recommended)
- RAM: ≥ 16 GB (48 GB recommended)
- Docker 19.03.9+, Docker Compose 1.25.1+
git clone https://github.com/dataelement/bisheng.git
cd bisheng/docker
docker compose -f docker-compose.yml -p bisheng up -d
Access via http://IP:3001. First registered user automatically becomes system admin.
Technical Architecture
Frontend (React + TypeScript)
↓
Backend (Python + FastAPI)
├── Workflow engine (LangChain/Langflow-based, customized)
├── Document parsing (proprietary OCR models)
├── Agent framework (AGL framework)
└── Model management (LLaMA-Factory integration)
↓
Data layer
├── Milvus (vector storage)
├── Elasticsearch (full-text search)
├── MySQL (metadata)
└── MinIO/S3 (file storage)
Acknowledgements: Built on LangChain, Langflow, Unstructured, and LLaMA-Factory, with extensive enterprise customization on top.
Links and Resources
- 🌟 GitHub: dataelement/bisheng
- 🌐 Website: bisheng.dataelem.com
- 📖 Documentation: github.com/dataelement/bisheng/wiki
Conclusion
BISHENG's core value is reducing the integration complexity of enterprise AI development.
Building an LLM application for an enterprise typically requires separately selecting a document parsing tool, RAG framework, workflow orchestration platform, agent framework, model management system, and permissions infrastructure — then connecting them all together. That integration process alone consumes substantial engineering resources. BISHENG consolidates all of these into a single platform with a unified data model, unified permissions, and unified monitoring.
High-precision document parsing is an underappreciated differentiator. In Chinese enterprise contexts, vast amounts of critical information still exists in scanned documents, handwritten forms, and stamped records. The quality of processing that data directly determines knowledge base usability. DataElem's 5+ years of proprietary training data represents a real advantage on this dimension.
The AGL framework and Human-in-the-Loop workflow reflect serious consideration of what makes enterprise scenarios different from general-purpose AI: not all decisions can be fully automated, and the right design preserves space for human intervention at critical junctures.
11.6k Stars and production use by Fortune 500 companies indicate this is not just a technical proof of concept — it's a product delivering measurable value.
Explore PrimeSkills — A marketplace for handpicked AI Agents and skills. Each is validated in real enterprise workflows, stripping away hype and keeping only what truly works.
Welcome to my Homepage for more useful insights and interesting products.
Top comments (0)