LangChain Intro by KeyMate.AI
- 1. FRAMEWORK LANGCHAIN “CHAIN” TOGETHER DIFFERENT COMPONENTS TO CREATE MORE ADVANCED USE CASES AROUND LLMS OZGUR OZKAN
- 2. MANAGING EXPECTATIONS Who am I? Startup Background Educational Background NEO AI Conversationalist Software Engineer Contractor @ Pollinate Founder @ Keymate.AI CTO in several tech startups one AI startup Reminis and one generative AI product in 2016 Masters in AI (Facial Recognition / Similarity Search) Bsc in CS My twitter bio set in 2019: I may continue to tweet even after I die thanks to future AI version of myself. Not dead, yet. ChatGPT rated this keynote 7/10
- 3. LANGCHAIN PHILOSOPHY 1.Connect a language model to other sources of data 2.Allow a language model to interact with its environment CLIK HERE NumPy and Pandas for LLMs, greatly increasing their usability and functionality Examples that leverages Langchain: AgentGPT, babyAGI Minecraft playing GPT4 (Voyager)***
- 4. LANGCHAIN COMPONENTS Schema: Text structure Model:OpenAI completion, text-in text-out or embedded Prompt templates Indexes and Document Loaders Memory: Long term and short term Chains: LLMChain Agents: which tools should be called or used Components Chains may consist of multiple components from several modules
- 5. Minecraft playing GPT-4 (Voyager) ReAct : REASONING AND ACTING IN LANGUAGE MODELS Plan-and-solve Chain-of-thought reasoning Had to develop my own langchain in iOS, tool using agent, chat history, decision prioritisation, message type handler ELEMENTS Focus on things on top Research Papers about prompting methods Langchain:LLM interaction framework but easy to adapt to another Programming Language and System
- 6. REIMPLEMENTING LANGCHAIN What if you have to use another PL (Not Python or TypeScript) ORCHESTRATOR
- 8. TEMPLATE
- 9. Can we run ReAct (Reason + Act) on ChatGPT Plus? It seems YES! we just need to find a way to inform LLM about thought patterns. Human structured/abstracted execution may not be the best option. (langchain) You as a human is just a tool for LLM. Langchain is already baked in ChatGPT.
- 11. REACT ON CHATGPT PLUS Utilize the "internetSearch" plugin and search for the Reasoning and Acting Chain of Thought framework. Generate reasoning traces and actions, then apply the ReAct Framework that you'll learn from the initial search. Subsequently, look up the latest research on the impact of climate change on biodiversity, and summarize the key findings. Too long, Didn't scan and read: Select the Keymate.AI search plugin and prompt in the following manner: 1. 2. Magic keyword is continue. Continues the loop. internetSearch Plugin (Keymate.AI) + 2 Extra tools of your choice
- 12. COMPONENTS CONT. LLMs (GPT4 , Hugging Face etc) In LLMs we trust all others manage time, resources, limits, autonomy Chat Models (Many built on top of GPT3) Embeddings (To store and search/retrieve big data) Toolkits (specialized agent for particular use case) Tools (agents can use to achieve certain tasks) Tokenizers (To count text size before passing to LLM) Document Loaders (Text document processing) Vectorstores (To store and index information to pass to agent) Agent Strategy (Prompt engineering / Research Papers ) The more general an agent is the less powerful in terms of task handling unless it has a very clever LLM. Good plugins can still work really well with ChatGPT. Interesting ideas: Add user based memory to your plugin. (auth and good vectorstore is needed) Add smart GPT4 based chains to your plugin ( time limit :( )
- 13. TOKEN LIMITS
- 15. TAKEAWAYS Can you observe the the thought reactions? When it awaits for input it wants to use you as a tool. Force trigger tools Using continue keyword Chain is derived from a dynamic state machine and it's endless You were part of the chain in ChatGPT and starting prompt Langchain is limited to two programming languages and limited platforms Build your own langchain. Good to grasp the concept. Amazing applications on top of langchain. Your language model can run structured or unstructured other models so that when right tools provided it can achieve anything. Personal opinion: Customised prompt templates and chains is better than using a framework, performance is limited with 30 seconds on the internet as sockets are short lived but LLM needs more chaining and time to execute sometimes. Using 10 tools at the same time is possible but not over the Web. Unstructured models should become structured and time-framed.
- 16. TAKEAWAYS Point 1: It's important to leverage human tool correctly. Users should be aware of using keywords and triggering specific tools when needed. ChatGPT is actually half-GPT, when you increase human UX and usability and prompt knowledge ChatGPT performs better. Human Tool can be enhanced to pass beyond the limits of chatgpt : Zero-shot, one-shot, few-shots learning concepts Tool triggering Usage of human memory to pass the context from one chat to another Transfer Learning Usage of Vector Databases for local memory, manual usage of vector databases to enhance human input. Naming entities and giving example to concepts. Forcing ChatGPT to review and rate itself. Although it sounds harsh to a human it makes chatGPT go beyond it's initial reasoning and pushes forward. User should know more about underlying agents: should I use the one that does constant self-critique or should I use the one that hallucinates a bit more. Time limits are the bottleneck of AI systems You need constant smart summarisation, divide n conquer techniques to overcome the issues.
- 17. TAKEAWAYS Specialized custom solutions work best; I suggest going bottom-up on expert systems, but structuring and abstracting things may not be beneficial. Langchain is abstraction and it leaks a lot; leakages cause time and money limits to hit early, and tasks usually either fail or require many iterations. You don't have to overoptimize on prompt engineering if LLM is clever. GPT4 learned how to apply ReAct framework with just a google search plugin. We need more clever LLMs. Good applications are very rare.
- 18. CONTACT ME Ozgur Ozkan ozgur.ozkan@keymate.ai EMAIL ADDRESS +447919236753 PHONE NUMBER WEBSITE keymate.ai LinkedIn
- 19. THANK YOU!!