Agentic topics
satya - 7/24/2024, 12:59:47 PM
Different names
- function calling
 - agents
 - planning
 - Take a number of actions in sequence
 - Agent takes more and more actions, the results of those actions are fed back to the LLM
 
satya - 7/24/2024, 2:32:48 PM
General observation
- Planning and reasoning is still not at the level it?s needed to be for real-world tasks.
 
satya - 7/24/2024, 2:40:44 PM
Issues
- prompt being passed into the LLM simply does not contain enough information for the LLM to make a reasonable decision
 - Lack of visibility to see intermediate data easily to evaluate what is not working
 - Test data sets that can be evaluated to understand the calculated flows because the underlying mechanism is brittle
 - randomness and unpredictability.
 
satya - 7/24/2024, 2:51:09 PM
Cognitive Architecture
- It is a term that is used to "describe" the "body" or "brain" of your "genai" application.
 - Particularly used in agentic systems as there is "reasoning" by LLM using a number of strategies including Prompt directing the LLM decisions to call which step next.
 - It is a way stemming from the desire to classify what kind of logic your have in your genai app and give it a name.
 
satya - 7/24/2024, 2:51:46 PM
Examples
- Straight forward question answer
 - basic RAG (no agentic)
 - Manual flows to control multiple calls to LLMs
 - Agentic RAG where "the routing logic is in the prompt" with multiple calls to the LLM and the APIs
 - Long term planning with goals
 
satya - 7/24/2024, 3:11:36 PM
Few Advanced Cognitive architectures
- Plan and Solve
 - Reflexion architecture
 - Or in combination with Domain specific "controlled" flows/architectures
 
satya - 7/24/2024, 3:19:38 PM
GenAI Agentic tool
ReACT:Search On Web
SayCan:Search On Web
ToolFormer:Search On Web
HuggingGPT:Search On Web
generative Agents:Search On Web
WebGPT:Search On Web
satya - 7/24/2024, 3:26:13 PM
Domain specific cognitive architecture: A manual (graph) flow in other words
- And rather than tell the LLM to do in the prompt
 - - where it may ignore it, not understand it, not have all the details -
 - they told the system to do it
 - by constructing a domain specific cognitive architecture.
 
satya - 7/24/2024, 10:01:14 PM
Essentials
- Takes LangChain as an example
 - Explains: Reason + Act
 - Introduces "Thought, Action, Observation"
 - A good article