Update README

README.md

@@ -1,26 +1,30 @@
-# Aggregated Chat Analysis for Personalised Teacher Assistance - Essay Feedback
+# 🤖 🤝 Multi LLM Agent Discovery, Structured Knowledge generation, and Collaboration
 
 As students use platforms like Khanmigo and MagicSchool.ai for tasks like getting feedback on their essays and "chat" with
-the tool to improve their learning, it will quickly become difficult for teachers to keep up with the volume of chat data.
+the tool to improve their learning, it will quickly become difficult for teachers to keep up with the volume of chat data to personalise instruction.
 
-This project aims to experiment with methods to aggregate chat data to provide teachers with data to improve instruction.
+In this repo we experiment with ways to store and aggregate structured knowledge from LLM Agents, and for agents to discover other agents by generating a thought to personalise their output.
 
-For the first experiment we will focus on a tool that gives feedback on essays.
+The [Agent Discovery](./agent-discovery-v1/agent-discovery.ipynb) notebook shows how student facing agents such as essay feedback providers can store knowledge about students that can be used by teacher facing agents such as a lesson plan generator to personalise the output.
 
-This aggregated feedback is then used to demonstrate the personalisation for a **Lesson Plan Generator Tool** that will be used by a Teacher.
+[ClassroomLM](https://classroom-lm.vercel.app/) is a WIP that will demonstrate the use of this experiment and here is the [frontend](https://github.com/13point5/classroom-lm) repo.
 
-[ClassroomLM](https://classroom-lm.vercel.app/) is a WIP and here is the [frontend](https://github.com/13point5/classroom-lm) repo.
+## Architecture
 
-## Data
+### 1. Multi Agent Discovery and Collaboration
 
-I used ChatGPT to generate the student essays and follow-up chat data. I used MagicSchool.ai's tool for providing feedback on essays for the AI messages.
+![multi agent architecture](./agent-architecture.png)
 
-## Architecture
+### 2. Knowledge Aggregation
 
-![Architecture](./architecture.png)
+![knowledge aggregation architecture](./architecture.png)
 
 I've used a simple `Evaluation` data model to analyse **ONLY** the AI messages. I know that this approach has drawbacks but I want to keep it simple for now to see if it's remotely useful.
 
+## Data
+
+I used ChatGPT to generate the student essays and follow-up chat data. I used MagicSchool.ai's tool for providing feedback on essays for the AI messages.
+
 ### Evaluation Data Model
 
 For each of the criteria below, we extract the following data:
@@ -62,7 +66,3 @@ For each of the criteria below, we extract the following data:
 ```bash
 poetry install
 ```
-
-## Usage
-
-The only experiment I have is in `experiment.ipynb` which is self-explanatory.

agent-discovery-v1/agent-discovery.ipynb

@@ -6,10 +6,35 @@
    "source": [
     "# Multi Agent Discovery, Knowledge generation, and Collaboration for Education\n",
     "\n",
-    "This notebook shows how AI Agents for Education that are used by students and teachers can:\n",
+    "This notebook shows how LLM based AI Agents for Education that are used by students and teachers can:\n",
     "1. Store aggregated structured knowledge about students\n",
     "2. Discover other agents by generating a `thought`\n",
-    "3. Collaborate by using the knowledge of other agents to `personalize` the output"
+    "3. Collaborate by using the knowledge of other agents to `personalize` the output\n",
+    "\n",
+    "## Architecture\n",
+    "![agent architecture](../agent-architecture.png)\n",
+    "\n",
+    "## Components:\n",
+    "### 1. `StudentAgent`\n",
+    "A `StudentAgent` is an agent designed to help students with specific expertise such as providing feedback on essays and a math tutor.\n",
+    "\n",
+    "These agents also use a `knowledge_class` to store aggregated structured knowledge about students through their interactions.\n",
+    "\n",
+    "A `knowledge_class` is a Pydantic model that describes the schema of knowledge to extract from the Agent's responses. This approach limits the quality of knowledge extracted but the purpose of this notebook is to demonstrate a possibility.\n",
+    "\n",
+    "Each `StudentAgent` also stores an automatically generated `knowledge_description` which describes the knowledge stored in plain text which will be used by other Agents to discover this agent.\n",
+    "\n",
+    "### 2. `StudentAgentStore`\n",
+    "\n",
+    "The `StudentAgentStore` is meant to resemble a database that stores all the agents and a vector database that stores the `knowledge_description`s for all agents for discovery.\n",
+    "\n",
+    "### 3. `TeacherAgent`\n",
+    "\n",
+    "A `TeacherAgent` is an agent designed to help teachers with tasks like generating a lesson plan or a worksheet.\n",
+    "\n",
+    "These agents `think` before responding so that they can determine if they need knowledge about students in a particular area to `personalise` the output.\n",
+    "\n",
+    "When they generate a thought describing the knowledge they need, they query the `StudentAgentStore` to get a relevant agent and ask the agent for the structured knowledge it has about students and uses it to assist teachers with their queries."
    ]
   },
   {