LLM in Enterprise Knowledge Management
Materials — open to everyone, no sign-in
Topic: LLM in Enterprise Knowledge Management
Presenter: Wenguang Wang
Additional Resources:
System Design Presentation - Databricks
Join Us on Wechat
Subscribe to Our YouTube channel https://commitway.com/eventyoutube
LLM: can understand knowledge across domains
KMS = (K x M ) ^S
K: knowledge
M: management
S: spreading
7 generations of knowledge management:
Knowledge database
Knowledge domain
Collaborative editing
Knowledge labeling
Knowledge graph
Big model / LLM
AGI
LLM includes GPT includes ChatGPT
ChatGTP brings hope for AGI (artificial general intelligence)
Reinforced learning / alignment
MOE
Intelligence system is much stronger than human intelligence
Current situation in knowledge management of companies in China
Cannot be accumulated
Cannot be found
Cannot be understood
Solutions
Accumulated
Visible
Values
Long term memory
Learn from mistakes
Grow employees
Improve productivity
Example:
Application in high speed rail
Solutions:
Automated analysis of knowledge
曹植 LLM
Knowledge graph
Architecture -> storage (graph database + vector database) -> application -> user interface
Knowledge management system
Knowledge architecture
Knowledge search
Knowledge community
Knowledge chain
Q&A
Knowledge delivery
Industries:
Medical
Automobile
Financial
Complexity
PDF including text and graphics
Needs deep knowledge of the industry
Knowledge delivery
Knowledge chain
Knowledge Q&A
RAG - retrieval augmented generation
Knowledge organization:
Knowledge search
Knowledge brain - full solution
Handling special situations
Not finding matched answers:
Can refuse to answer
Or can use LLM to answer, but label the answer may contain hallucination
Lots of challenges
Hallucination
Refresh of knowledge
Multi-tier solution
Knowledge enhancement: RAG, knowledge graph, plugin, agent
Model ability
Other challenges
Language is only part of the knowledge
Non language knowledge
Hard to reach human level intelligence by only relying on language
AGI
Multi-model
Induction
Use of tools
AGI challenges
Software future
Engineer will live
Software will die
Medium company?
Difficult to adopt LLM/knowledge management system
Can use SaaS based LLM, such as GPT/co-pilot
Customer case?
Electrical utility company - troubleshooting
GM - troubleshooting
Challenges:
Lack of cooperation between different departments
knowledge/documentation incomplete
Future for China/US
In the short run the gap will increase; in the long run the gap will shrink
Limitation of GPU
But in the future China may overcome the lack of GPU
Lots of opportunity in China
Very big market - can specialize in one vertical (Steel, airline,etc)
Price competition
Price war is caused by competition
国内很卷
Startup ecosystem
In china: Difficult to be funded. Lots of market
Easier in silicon valley
Specialty in 达观
Graph + LLM
Troubleshooting scenario
Market is big enough
Cannot take all market - expect competition
How to update knowledge
Can ingest from other systems
Can upload
Can write adapters
100 people for development
200 people customization/deployment
Knowledge graph?
GensGraph
Nebula graph
How do the UIs integrate together?
If they have an existing system - usually don’t buy a new system
Integrate
Replace
RAG?
Usually not effective
Usually customize code / dirty work - lots of scenarios
Integrate with knowledge graph