This is based on the DeepLearning.AI, AI Agents in LangGraph course. In this repository I have converted all of the examples from Python to JavaScript.
Agentic use and agentic search have moved on a lot recently. The sort of workflow discussed is one in which a request is made, actions to find information are taken and the results are written up. This is an iterative process until the goal is reached. This involves:
- Planning
- Tool use
- Reflection
- Multi-agent communication
- Memory Langchain offers many of these elements, but has recently improved agent support with Cyclic Graphs.
ReAct (Reason + Act)
====================
Actions
------------- -------------
| | | |
| v | v
Reasoning LLM Env
Traces
^ | ^ |
| | | |
------------- -------------
Observations
ReAct: Synergizing reasoning and acting in language models
Self-Refine
===========
Input
------------- | -------------
| | | | |
v | v | v
Feedback Model M Refine
| ^ ^ |
| | | |
------------- -------------
Use M to get Use M to refine
feedback on its its previous output
own output given the feedback
Self-Refine: Iterative refinement with self-feedback
AlphaCodium
===========
PRE-PROCESSING CODE ITERATIONS
------------------------------------------------------- --------------------------------------
_____________
| |
v |
Input - Problem Generate Rank Iterate on | Iterate on
Description + Possible ---------> Solutions Public Tests ---------> AI Tests <-
Public Tests Solutions | ^ | |
| ^ | | | |
| | | |_____ |-------
v | v | | v
Problem Public Tests Generate Initial | Final
Reflection ---------> Reasoning Additional ---------> Code <- Solution
AI Tests Solution
Code engineering with AlphaCodium: From prompt engineering to flow engineering
Langchain has been extended to include LangGraph to support these.
- Build an Agent from Scratch
- LangGraph Components
- Agentic Search Tools
- Persistence and Streaming
- Human in the loop
- Essay Writer
- LangChain Resources
If you want to try these out you will first need to setup your own ChatGPT secret key in your local environment. Here is how you get a key. Once you have this put it in a local (server side) environment variable. For example in Mac OS, assuming you are using zsh, append the following to the file .zshenv in you own home directory:
export OPENAI_API_KEY='your_secret_key_value'
When you restart the shell or your machine the environment variable OPENAI_API_KEY will be in place.
This tutorial also uses the Tavily search tool. To use this you need to sign up for an API key and also export that API key in the same way we exported the OpenAPI key for ChatGPT. In this case you need to add the line:
export TAVILY_API_KEY='your_tavily_api_key'
Add this into your environment settings for the shell you are using.
Before trying any of the exercises don't forget to run npm install in the ./agents-langraph directory to install the Node modules needed.
In each subdirectory you will find a *.js or a *.ts file and, sometimes, some supporting files. Each file contains multiple prompts.
In most cases the initial exercise is ready to run and the other exercises are commented out using the \* ... *\ comment markers. In these cases the commented code blocks will have their own calls to the LLM. If you uncomment these blocks then be sure to comment out the last to calls above while you run that exercise, it will reduce run time and costs.
There are some architectures it is worth knpwing about but which are not covered by this course. These include:
Multiple agents working on the same state. All work on the same shared state.
-----------------------------------
| User |
| Input: Generate a chart of ave. |
| temp. in Alaska over last decade. |
-----------------------------------
|
First go to researcher
|
v
-------------------------- ----------------------- -----------------
| Researcher |--------- message ------->| Router |<-------- message --------| Chart Generator |
| Call a 'search' function | | (If statements based) | | |
| or FINISH |<--- If 'continue' and ---| on agent output) |--- If 'continue' and --->| Code execution |
-------------------------- state.sender == ----------------------- state.sender == -----------------
^ 'chart_generator' | 'researcher' ^
| | |
If state.sender == If function is called If state.sender ==
'researcher' | 'chart_generator'
| v |
| ----------- |
--------------------------------------------| call_tool |-----------------------------------------
-----------
In this case the state may not be shared, each Agent may be a separate graph. The supervisor can use a more powerful LLM.
-------
| User |
-------
| ^
| |
v |
-----------------------------------
| Supervisor |
-----------------------------------
^ | | ^ ^ |
| route route | | route
| | | | | |
| v v | | v
--------- --------- ---------
| Agent 1 | | Agent 2 | | Agent 3 |
--------- --------- ---------
See the Codium model above, this is essentially a pipeline with some noted that loop. The plan and execute style flow is a simpler example of this:
------
---------->| Plan |
| ------
| |
| 2 generate tasks
1 user v
request ----------- -----------------
| | Task list | | |
------ | * ~~~~~~~ | ------------------- v
| User | | * ~~~~~~~ |-- 3 exec tasks -->| Single-task-agent | Loop to solve task <--> TOOL
------ | * ~~~~~~~ | ------------------- |
^ | * ~~~~~~~ | | ^ |
| ----------- | -----------------
| ^ 4 update state
| 5b replan more tasks with task results
5a respond | |
to user -------- |
----------| Replan |<-------------------------------
--------
This approach repeats until solved:
- Select Node
- Generate new candidates
- Act, reflect and score
- Back propagate (update parents) This makes great use of persistence to jump back to previous times when needed.
---------- --------- ---------- --------- ---------- ---------
| Generate | | Reflect | | Generate | | Reflect | | Generate | | Reflect |
---------- ^--------- ^---------- ^--------- ^---------- ^---------
| / | / | / | | / / | | |
v / v / v / | | / / | v |
---------- / --------- / ---------- / -|-----|- / ---------- / |------|-
| %Act | | %Act .3 | | %Act .3 | | %Act .6 | | %Act .6 | ||%Act .7 |
---------- --------- ---------- -|-----|- ---------- |------|-
| | | ^ ^ | | | ^ |
v v v/ \v v | / |
--- --- ----- ----- ----- ----- --|-- --|--
| % | | % | | %.8 || %.4 || %.8 || %.4 || %.9 || %.4 |
--- --- ----- ----- ----- ----- --|-- --|--
| | | ^^ |
v v v | \ v
--- --- ----- -----
| % || % | | %.2 || % 1 |
--- --- ----- -----