Fimeg/letta-server
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

fix: various fixes to python client and add tutorial notebooks (#1377)

Browse Source
This commit is contained in:
Sarah Wooders
2024-06-09 15:32:20 -07:00
committed by GitHub
parent de7de7e912
commit 51f06fdae2
26 changed files with 931 additions and 82 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
.github/workflows/tests.yml vendored
View File

@@ -68,7 +68,7 @@ jobs:
MEMGPT_SERVER_PASS: test_server_token
PYTHONPATH: ${{ github.workspace }}:${{ env.PYTHONPATH }}
run: |
poetry run pytest -s -vv -k "not test_quickstart and not test_endpoints and not test_storage and not test_server and not test_openai_client" tests
poetry run pytest -s -vv -k "not test_concurrent_connections.py and not test_quickstart and not test_endpoints and not test_storage and not test_server and not test_openai_client" tests
- name: Run storage tests
env:

2
Dockerfile
View File

@@ -16,7 +16,7 @@ RUN poetry lock --no-update
RUN if [ "$MEMGPT_ENVIRONMENT" = "DEVELOPMENT" ] ; then \
poetry install --no-root -E "postgres server dev autogen" ; \
else \
poetry install --without dev --no-root -E "postgres server" && \
poetry install --without dev --without local --no-root -E "postgres server" && \
rm -rf $POETRY_CACHE_DIR ; \
fi

1
dev-compose.yaml
View File

@@ -35,6 +35,7 @@ services:
- MEMGPT_PG_HOST=pgvector_db
- MEMGPT_PG_PORT=5432
- OPENAI_API_KEY=${OPENAI_API_KEY}
- SERPAPI_API_KEY=${SERPAPI_API_KEY}
volumes:
- ./configs/server_config.yaml:/root/.memgpt/config # config file
# ~/.memgpt/credentials:/root/.memgpt/credentials # credentials file

BIN
examples/tutorials/dev_portal_agent_chat.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 134 KiB

BIN
examples/tutorials/dev_portal_memory.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 266 KiB

BIN
examples/tutorials/dev_portal_tools.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 488 KiB

BIN
examples/tutorials/developer_portal_login.png Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 100 KiB

239
examples/tutorials/local-python-client.ipynb Normal file
View File

@@ -0,0 +1,239 @@
{
"cells": [
{
"cell_type": "markdown",
"id": "c015b59e-1187-4d45-b2af-7b4c5a9512e1",
"metadata": {},
"source": [
"# MemGPT Python Client \n",
"Welcome to the MemGPT tutorial! In this tutorial, we'll go through how to create a basic user-client for MemGPT and create a custom agent with long term memory. \n",
"\n",
"MemGPT runs *agents-as-a-service*, so agents can run independently on a server. For this tutorial, we will run a local version of the client which does not require a server, but still allows you to see some of MemGPT's capabilities. "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "a34fe313-f63e-4f36-9142-f681431bbb91",
"metadata": {},
"outputs": [],
"source": [
"!pip install git+https://github.com/cpacker/MemGPT.git@tutorials"
]
},
{
"cell_type": "markdown",
"id": "191c1cf1-03e6-411a-8409-003caa8530f5",
"metadata": {},
"source": [
"### Setup your OpenAI API key "
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "23091690-bc50-4fbc-b48d-50b639453e36",
"metadata": {},
"outputs": [],
"source": [
"import os \n",
"\n",
"os.environ[\"OPENAI_API_KEY\"] = \"sk-...\""
]
},
{
"cell_type": "markdown",
"id": "f20ad6c7-9066-45e0-88ac-40920c83cc39",
"metadata": {},
"source": [
"## Part 1: Connecting to the MemGPT Client \n",
"\n",
"We create a local client which creates a quickstart configuration for OpenAI using the provided `OPENAI_API_KEY`. "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "9b0871a0-42af-4573-a8ba-efb4fe7e5e5a",
"metadata": {},
"outputs": [],
"source": [
"from memgpt.client.client import LocalClient\n",
"\n",
"client = LocalClient(quickstart_option=\"openai\") "
]
},
{
"cell_type": "markdown",
"id": "40666896-0fa2-465e-b51b-57719de30542",
"metadata": {},
"source": [
"## Part 2: Create an agent \n",
"We'll first start with creating a basic MemGPT agent. "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "fb90f12b-acd7-4877-81e8-0e7b9eb4bd9b",
"metadata": {},
"outputs": [],
"source": [
"basic_agent = client.create_agent(\n",
" name=\"basic_agent\", \n",
")\n",
"print(f\"Created agent: {basic_agent.name}\")"
]
},
{
"cell_type": "markdown",
"id": "94d14102-3ef8-40fe-b32e-c77d0b8df311",
"metadata": {},
"source": [
"We can now send messages from the user to the agent by specifying the `agent_id`: "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "3cbfef36-76f0-4f0b-990a-5d8409a676d7",
"metadata": {},
"outputs": [],
"source": [
"from memgpt.client.utils import pprint \n",
"\n",
"response = client.user_message(agent_id=basic_agent.id, message=\"hello\") \n",
"pprint(response.messages)"
]
},
{
"cell_type": "markdown",
"id": "b24d048e-f3cc-4830-aaa2-5e590d652bd9",
"metadata": {},
"source": [
"### Adding Personalization\n",
"We can now create a more customized agent, but specifying a custom `human` and `persona` field. \n",
"* The *human* specifies the personalization information about the user interacting with the agent \n",
"* The *persona* specifies the behavior and personality of the event\n",
"\n",
"What makes MemGPT unique is that the starting *persona* and *human* can change over time as the agent gains new information, enabling it to have evolving memory. We'll see an example of this later in the tutorial."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "3ec35979-9102-4ea7-926e-ea7ccd501ceb",
"metadata": {},
"outputs": [],
"source": [
"# TODO: feel free to change the human and person to what you'd like \n",
"persona = \\\n",
"\"\"\"\n",
"You are a friendly and helpful agent!\n",
"\"\"\"\n",
"\n",
"human = \\\n",
"\"\"\"\n",
"I am an Accenture consultant with many specializations. My name is Sarah.\n",
"\"\"\"\n",
"\n",
"custom_agent = client.create_agent(\n",
" name=\"custom_agent\", \n",
" human=human, \n",
" persona=persona\n",
")"
]
},
{
"cell_type": "markdown",
"id": "63a9a61b-58c9-4d09-a4f7-48233c72c340",
"metadata": {},
"source": [
"### Viewing memory \n",
"You can access the agent's memories through the client. There are two type of memory, *core* and *archival* memory: \n",
"1. Core memory stores short-term memories in the LLM's context \n",
"2. Archival memory stores long term memories in a vector database\n",
"\n",
"Core memory is divided into a \"human\" and \"persona\" section. You can see the agent's memories about the human below: "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "b0d1840a-05ee-47c1-b5f5-89faafd96e7c",
"metadata": {},
"outputs": [],
"source": [
"print(client.get_agent_memory(agent_id=custom_agent.id)[\"core_memory\"][\"human\"])"
]
},
{
"cell_type": "markdown",
"id": "95c8a058-5d67-45b7-814b-38bb67c9acf3",
"metadata": {},
"source": [
"### Evolving memory \n",
"MemGPT agents have long term memory, and can evolve what they store in their memory over time. In the example below, we make a correction to the previously provided information. See how the agent processes this new information. "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "7e58e685-579e-4a0d-bba7-41976ea7f469",
"metadata": {},
"outputs": [],
"source": [
"response = client.user_message(agent_id=custom_agent.id, message=\"Actually, my name is Charles\") \n",
"pprint(response.messages)"
]
},
{
"cell_type": "markdown",
"id": "af2a2dd6-925e-49b2-ab01-bf837f33b26c",
"metadata": {},
"source": [
"Now lets see what the agent's memory looks like again: "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "41ef4aaa-4a48-44bb-8944-855f30725d6d",
"metadata": {},
"outputs": [],
"source": [
"print(client.get_agent_memory(agent_id=custom_agent.id)[\"core_memory\"][\"human\"])"
]
},
{
"cell_type": "markdown",
"id": "66da949b-1084-4b87-b77c-6cbd4a822b34",
"metadata": {},
"source": [
"## 🎉 Congrats, you're done with day 1 of MemGPT! \n",
"For day 2, we'll go over how to connect *data sources* to MemGPT to run RAG agents. "
]
}
],
"metadata": {
"kernelspec": {
"display_name": "memgpt",
"language": "python",
"name": "memgpt"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.2"
}
},
"nbformat": 4,
"nbformat_minor": 5
}

50
examples/tutorials/memgpt-admin-client.ipynb Normal file
View File

@@ -0,0 +1,50 @@
{
"cells": [
{
"cell_type": "code",
"execution_count": null,
"id": "fb13c7bc-fbb4-4ccd-897c-08995db258e8",
"metadata": {},
"outputs": [],
"source": [
"from memgpt import Admin \n",
"\n",
"base_url=\"memgpt.localhost\"\n",
"token=\"memgptadmin\" \n",
"\n",
"admin_client = Admin(base_url=base_url, token=\"memgptadmin\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "984b8249-a3f7-40d1-9691-4d128f9a90ff",
"metadata": {},
"outputs": [],
"source": [
"user = admin_client.create_user()"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "memgpt",
"language": "python",
"name": "memgpt"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.2"
}
},
"nbformat": 4,
"nbformat_minor": 5
}

BIN
examples/tutorials/memgpt_paper.pdf Normal file
View File

Binary file not shown.

125
examples/tutorials/memgpt_rag_agent.ipynb Normal file
View File

@@ -0,0 +1,125 @@
{
"cells": [
{
"cell_type": "markdown",
"id": "64fa991c-98e5-4be0-a838-06a4617d8be3",
"metadata": {},
"source": [
"## Part 4: Adding external data \n",
"In addition to short term, in-context memories, MemGPT agents also have a long term memory store called *archival memory*. We can enable agents to leverage external data (e.g. PDF files, database records, etc.) by inserting data into archival memory. In this example, we'll show how to load the MemGPT paper a *source*, which defines a set of data that can be attached to agents. "
]
},
{
"cell_type": "markdown",
"id": "c61ac9c3-cbea-47a5-a6a4-4133ffe5984e",
"metadata": {},
"source": [
"We first download a PDF file, the MemGPT paper: "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "f89e9156-3d2d-4ce6-b5e9-aeb4cdfd5657",
"metadata": {},
"outputs": [],
"source": [
"import requests\n",
"\n",
"url = \"https://arxiv.org/pdf/2310.08560\"\n",
"response = requests.get(url)\n",
"filename = \"memgpt_paper.pdf\"\n",
"\n",
"with open(filename, 'wb') as f:\n",
" f.write(response.content)"
]
},
{
"cell_type": "markdown",
"id": "bcfe3a48-cdb0-4843-9599-623753eb61b9",
"metadata": {},
"source": [
"Next, we create a MemGPT source to load data into: "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "7ccf21fb-5862-42c2-96ca-63e0ba2f48b5",
"metadata": {},
"outputs": [],
"source": [
"memgpt_paper = client.create_source(\n",
" name=\"memgpt_paper\", \n",
")"
]
},
{
"cell_type": "markdown",
"id": "f114bf0b-6a25-4dbf-9c2c-59271d46ebba",
"metadata": {},
"source": [
"Now that we have a source, we can load files into the source. Loading the file will take a bit of time, since the file needs to be parsed and stored as *embeddings* using an embedding model. The loading function returns a *job* which can be pinged for a status. "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "6fe624eb-bf08-4267-a849-06103c1ad5b6",
"metadata": {},
"outputs": [],
"source": [
"job = client.load_file_into_source(filename=filename, source_id=memgpt_paper.id)\n",
"job"
]
},
{
"cell_type": "markdown",
"id": "27ce13f5-d878-406d-9a5f-7e2335f2ef0d",
"metadata": {},
"source": [
"### Attaching data to an agent \n",
"To allow an agent to access data in a source, we need to *attach* it to the agent. This will load the source's data into the agent's archival memory. "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "5be91571-87ee-411a-8e79-25c56c414360",
"metadata": {},
"outputs": [],
"source": [
"client.attach_source_to_agent(source_id=memgpt_paper.id, agent_id=basic_agent.id)\n",
"# TODO: add system message saying that file has been attached \n",
"\n",
"from pprint import pprint\n",
"\n",
"# TODO: do soemthing accenture related \n",
"# TODO: brag about query rewriting -- hyde paper \n",
"response = client.user_message(agent_id=basic_agent.id, message=\"what is core memory? search your archival memory.\") \n",
"pprint(response.messages)"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "memgpt",
"language": "python",
"name": "memgpt"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.2"
}
},
"nbformat": 4,
"nbformat_minor": 5
}

319
examples/tutorials/python-client.ipynb Normal file
View File

@@ -0,0 +1,319 @@
{
"cells": [
{
"cell_type": "markdown",
"id": "6d3806ac-38f3-4999-bbed-953037bd0fd9",
"metadata": {},
"source": [
"# MemGPT Python Client \n",
"Welcome to the MemGPT tutorial! In this tutorial, we'll go through how to create a basic user-client for MemGPT and create a custom agent with long term memory. \n",
"\n",
"MemGPT runs *agents-as-a-service*, so agents can run independently on a server. For this tutorial, we will be connecting to an existing MemGPT server via the Python client and the UI console. If you don't have a running server, see the [documentation](https://memgpt.readme.io/docs/running-a-memgpt-server) for instructions on how to create one. "
]
},
{
"cell_type": "markdown",
"id": "7c0b6d6b-dbe6-412b-b129-6d7eb7d626a3",
"metadata": {},
"source": [
"## Part 0: Install MemGPT "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "481d0976-d26b-46d2-ba74-8f2bb5556387",
"metadata": {},
"outputs": [],
"source": [
"!pip install git+https://github.com/cpacker/MemGPT.git@tutorials"
]
},
{
"cell_type": "markdown",
"id": "a0484348-f7b2-48e3-9a2f-7d6495ef76e3",
"metadata": {},
"source": [
"## Part 1: Connecting to the MemGPT Client \n",
"\n",
"The MemGPT client connects to a running MemGPT service, specified by `base_url`. The client corresponds to a *single-user* (you), so requires an authentication token to let the service know who you are. \n",
"\n"
]
},
{
"cell_type": "code",
"execution_count": 19,
"id": "53ae2e1b-ad22-43c2-b3d8-92d591be8840",
"metadata": {},
"outputs": [],
"source": [
"from memgpt import create_client\n",
"\n",
"base_url = \"http://35.238.125.250:8083\"\n",
"\n",
"# TODO: replace with your token \n",
"my_token = \"sk-...\" \n",
"\n",
"client = create_client(base_url=base_url, token=my_token) "
]
},
{
"cell_type": "markdown",
"id": "3c5c8651-e8aa-4423-b2b8-284bf6a01577",
"metadata": {},
"source": [
"### Viewing the developer portal \n",
"MemGPT provides a portal interface for viewing and interacting with agents, data sources, tools, and more. You can enter `http://35.238.125.250:8083` into your browser to load the developer portal, and enter in `my_token` to log in. \n",
"\n",
"<img src=\"./developer_portal_login.png\" width=\"800\">"
]
},
{
"cell_type": "markdown",
"id": "66e47b34-5feb-4660-85f0-14b5ee7f62b9",
"metadata": {},
"source": [
"## Part 2: Create an agent \n",
"We'll first start with creating a basic MemGPT agent. "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "24745606-b0fb-4157-a5cd-82fd0c26711f",
"metadata": {},
"outputs": [],
"source": [
"basic_agent = client.create_agent(\n",
" name=\"basic_agent\", \n",
")\n",
"print(f\"Created agent: {basic_agent.name}\")"
]
},
{
"cell_type": "markdown",
"id": "fcfb0d7b-b260-4bc0-8db2-c65f40e4afd5",
"metadata": {},
"source": [
"We can now send messages from the user to the agent by specifying the `agent_id`: "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "a37bc9aa-4efb-4b4d-a6ce-f02505cb3240",
"metadata": {},
"outputs": [],
"source": [
"from memgpt.client.utils import pprint \n",
"\n",
"response = client.user_message(agent_id=basic_agent.id, message=\"hello\") \n",
"pprint(response.messages)"
]
},
{
"cell_type": "markdown",
"id": "9803140c-2b9d-426b-8812-9295806eb312",
"metadata": {},
"source": [
"### Chatting in the developer portal \n",
"You can also chat with the agent inside of the developer portal. Try clicking the chat button in the agent view. \n",
"\n",
"<img src=\"./dev_portal_agent_chat.png\" width=\"800\">"
]
},
{
"cell_type": "markdown",
"id": "99ae20ec-e92e-4480-a652-b4aea28a6199",
"metadata": {},
"source": [
"### Adding Personalization\n",
"We can now create a more customized agent, but specifying a custom `human` and `persona` field. \n",
"* The *human* specifies the personalization information about the user interacting with the agent \n",
"* The *persona* specifies the behavior and personality of the event\n",
"\n",
"What makes MemGPT unique is that the starting *persona* and *human* can change over time as the agent gains new information, enabling it to have evolving memory. We'll see an example of this later in the tutorial."
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "c0876410-4d70-490d-a798-39938b5ce941",
"metadata": {},
"outputs": [],
"source": [
"# TODO: feel free to change the human and person to what you'd like \n",
"persona = \\\n",
"\"\"\"\n",
"You are a friendly and helpful agent!\n",
"\"\"\"\n",
"\n",
"human = \\\n",
"\"\"\"\n",
"I am an Accenture consultant with many specializations. My name is Sarah.\n",
"\"\"\"\n",
"\n",
"custom_agent = client.create_agent(\n",
" name=\"custom_agent\", \n",
" human=human, \n",
" persona=persona\n",
")"
]
},
{
"cell_type": "markdown",
"id": "21293857-80e4-46e4-b628-3912fad038e9",
"metadata": {},
"source": [
"### Viewing memory \n",
"You can view and edit the agent's memory inside of the developer console. There are two type of memory, *core* and *archival* memory: \n",
"1. Core memory stores short-term memories in the LLM's context \n",
"2. Archival memory stores long term memories in a vector database\n",
"\n",
"In this example, we'll look at how the agent can modify its core memory with new information. To see the agent's memory, click the \"Core Memory\" section on the developer console. \n",
"\n",
"<img src=\"./dev_portal_memory.png\" width=\"800\">"
]
},
{
"cell_type": "markdown",
"id": "d8fa13eb-ce4b-4e4f-81b6-9d6ef6fa67c2",
"metadata": {},
"source": [
"### Referencing memory \n",
"MemGPT agents can customize their responses based on what memories they have stored. Try asking a question that related to the human and persona you provided. "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "fddbefe5-3b94-4a08-aa50-d80fb581c747",
"metadata": {},
"outputs": [],
"source": [
"response = client.user_message(agent_id=custom_agent.id, message=\"what do I work as?\") \n",
"pprint(response.messages)"
]
},
{
"cell_type": "markdown",
"id": "30497119-e208-4a4e-b482-e7cfff346263",
"metadata": {},
"source": [
"### Evolving memory \n",
"MemGPT agents have long term memory, and can evolve what they store in their memory over time. In the example below, we make a correction to the previously provided information. See how the agent processes this new information. "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "679fa708-20ee-4e75-9222-b476f126bc6f",
"metadata": {},
"outputs": [],
"source": [
"response = client.user_message(agent_id=custom_agent.id, message=\"Actually, my name is Charles\") \n",
"pprint(response.messages)"
]
},
{
"cell_type": "markdown",
"id": "686ac5a3-be63-4afd-97ae-b7d05219dd60",
"metadata": {},
"source": [
"Now, look back at the developer portal and at the agent's *core memory*. Do you see a change in the *human* section of the memory? "
]
},
{
"cell_type": "markdown",
"id": "878d2f49-a5a6-4483-9f69-7436bcf00cfb",
"metadata": {},
"source": [
"## Part 3: Adding Tools \n",
"MemGPT agents can be connected to custom tools. Currently, tools must be created by service administrators. However, you can add additional tools provided by the service administrator to the agent you create. "
]
},
{
"cell_type": "markdown",
"id": "35785d36-2674-4a00-937b-4c747e0fb6bf",
"metadata": {},
"source": [
"### View Available Tools "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "c307a6f7-276b-49f5-8d3d-48aaaea221a7",
"metadata": {},
"outputs": [],
"source": [
"tools = client.list_tools().tools\n",
"for tool in tools: \n",
" print(f\"Tool: {tool.name} - {tool.json_schema['description']}\")"
]
},
{
"cell_type": "markdown",
"id": "318d19dc-b9dd-448c-ab5c-9c9311d21fad",
"metadata": {},
"source": [
"### Create a tool using agent in the developer portal \n",
"Create an agent in the developer portal and toggle additional tools you want the agent to use. We recommend modifying the *persona* to notify the agent that it should be using the tools for certain tasks. \n",
"\n",
"\n",
"<img src=\"./dev_portal_tools.png\" width=\"800\">"
]
},
{
"cell_type": "markdown",
"id": "aecdaa70-861a-43d5-b006-fecd90a8ed19",
"metadata": {},
"source": [
"## Part 4: Cleanup (optional) \n",
"You can cleanup the agents you creating the following command to delete your agents: "
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "1320d9c9-170b-48a8-b5e8-70737b1a8aac",
"metadata": {},
"outputs": [],
"source": [
"for agent in client.list_agents().agents: \n",
" client.delete_agent(agent[\"id\"])\n",
" print(f\"Deleted agent {agent['name']} with ID {agent['id']}\")"
]
},
{
"cell_type": "markdown",
"id": "510675a8-22bc-4f9f-9c79-91e2ffa9caf9",
"metadata": {},
"source": [
"## 🎉 Congrats, you're done with day 1 of MemGPT! \n",
"For day 2, we'll go over how to connect *data sources* to MemGPT to run RAG agents. "
]
}
],
"metadata": {
"kernelspec": {
"display_name": "memgpt",
"language": "python",
"name": "memgpt"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.2"
}
},
"nbformat": 4,
"nbformat_minor": 5
}

7
memgpt/agent.py
View File

@@ -86,9 +86,14 @@ def link_functions(function_schemas: list):
raise ValueError(f"While loading agent.state.functions encountered a bad function schema object with no name:\n{f_schema}")
linked_function = available_functions.get(f_name)
if linked_function is None:
raise ValueError(
# raise ValueError(
# f"Function '{f_name}' was specified in agent.state.functions, but is not in function library:\n{available_functions.keys()}"
# )
print(
f"Function '{f_name}' was specified in agent.state.functions, but is not in function library:\n{available_functions.keys()}"
)
continue
# Once we find a matching function, make sure the schema is identical
if json.dumps(f_schema, ensure_ascii=JSON_ENSURE_ASCII) != json.dumps(
linked_function["json_schema"], ensure_ascii=JSON_ENSURE_ASCII

9
memgpt/client/admin.py
View File

@@ -63,7 +63,6 @@ class Admin:
if response.status_code != 200:
raise HTTPError(response.json())
response_json = response.json()
print(response_json)
return CreateUserResponse(**response_json)
def delete_user(self, user_id: uuid.UUID):
@@ -89,23 +88,23 @@ class Admin:
"""Add a tool implemented in a file path"""
source_code = open(file_path, "r", encoding="utf-8").read()
data = {"name": name, "source_code": source_code, "source_type": source_type, "tags": tags}
response = requests.post(f"{self.base_url}/api/tools", json=data, headers=self.headers)
response = requests.post(f"{self.base_url}/admin/tools", json=data, headers=self.headers)
if response.status_code != 200:
raise ValueError(f"Failed to create tool: {response.text}")
return ToolModel(**response.json())
def list_tools(self) -> ListToolsResponse:
response = requests.get(f"{self.base_url}/api/tools", headers=self.headers)
response = requests.get(f"{self.base_url}/admin/tools", headers=self.headers)
return ListToolsResponse(**response.json())
def delete_tool(self, name: str):
response = requests.delete(f"{self.base_url}/api/tools/{name}", headers=self.headers)
response = requests.delete(f"{self.base_url}/admin/tools/{name}", headers=self.headers)
if response.status_code != 200:
raise ValueError(f"Failed to delete tool: {response.text}")
return response.json()
def get_tool(self, name: str):
response = requests.get(f"{self.base_url}/api/tools/{name}", headers=self.headers)
response = requests.get(f"{self.base_url}/admin/tools/{name}", headers=self.headers)
if response.status_code == 404:
return None
elif response.status_code != 200:

3
memgpt/client/client.py
View File

@@ -239,7 +239,6 @@ class RESTClient(AbstractClient):
super().__init__(debug=debug)
self.base_url = base_url
self.headers = {"accept": "application/json", "authorization": f"Bearer {token}"}
self.token = token
# agents
@@ -696,7 +695,7 @@ class LocalClient(AbstractClient):
if self.auto_save:
self.save()
else:
return self.interface.to_list()
return UserMessageResponse(messages=self.interface.to_list())
def run_command(self, agent_id: str, command: str) -> Union[str, None]:
self.interface.clear()

61
memgpt/client/utils.py Normal file
View File

@@ -0,0 +1,61 @@
from datetime import datetime
from IPython.display import HTML, display
def pprint(messages):
"""Utility function for pretty-printing the output of client.send_message in notebooks"""
css_styles = """
<style>
.terminal {
background-color: #002b36;
color: #839496;
font-family: 'Courier New', Courier, monospace;
padding: 10px;
border-radius: 5px;
}
.terminal strong {
color: #b58900;
}
.terminal .function-return {
color: #2aa198;
}
.terminal .internal-monologue {
color: #d33682;
}
.terminal .function-call {
color: #2aa198;
}
.terminal .assistant-message {
color: #859900;
}
.terminal pre {
color: #839496;
}
</style>
"""
html_content = css_styles + "<div class='terminal'>"
for message in messages:
date_str = message["date"]
date_formatted = datetime.fromisoformat(date_str.replace("Z", "+00:00")).strftime("%Y-%m-%d %H:%M:%S")
if "function_return" in message:
return_string = message["function_return"]
return_status = message["status"]
html_content += f"<p><strong>🛠️ [{date_formatted}] Function Return ({return_status}):</strong></p>"
html_content += f"<p class='function-return'>{return_string}</p>"
elif "internal_monologue" in message:
html_content += f"<p><strong>💭 [{date_formatted}] Internal Monologue:</strong></p>"
html_content += f"<p class='internal-monologue'>{message['internal_monologue']}</p>"
elif "function_call" in message:
html_content += f"<p><strong>🛠️ [[{date_formatted}] Function Call:</strong></p>"
html_content += f"<p class='function-call'>{message['function_call']}</p>"
elif "assistant_message" in message:
html_content += f"<p><strong>🤖 [{date_formatted}] Assistant Message:</strong></p>"
html_content += f"<p class='assistant-message'>{message['assistant_message']}</p>"
html_content += "<br>"
html_content += "</div>"
display(HTML(html_content))

8
memgpt/metadata.py
View File

@@ -87,7 +87,6 @@ class LLMConfigColumn(TypeDecorator):
return value
def process_result_value(self, value, dialect):
# print("GET VALUE", value)
if value:
return LLMConfig(**value)
return value
@@ -400,10 +399,7 @@ class MetadataStore:
@enforce_types
def get_api_key(self, api_key: str) -> Optional[Token]:
with self.session_maker() as session:
print("getting api key", api_key)
print([r.token for r in self.get_all_api_keys_for_user(user_id=uuid.UUID(int=0))])
results = session.query(TokenModel).filter(TokenModel.token == api_key).all()
print("results", [r.token for r in results])
if len(results) == 0:
return None
assert len(results) == 1, f"Expected 1 result, got {len(results)}" # should only be one result
@@ -414,18 +410,15 @@ class MetadataStore:
with self.session_maker() as session:
results = session.query(TokenModel).filter(TokenModel.user_id == user_id).all()
tokens = [r.to_record() for r in results]
print([t.token for t in tokens])
return tokens
@enforce_types
def get_user_from_api_key(self, api_key: str) -> Optional[User]:
"""Get the user associated with a given API key"""
token = self.get_api_key(api_key=api_key)
print("got api key", token.token, token is None)
if token is None:
raise ValueError(f"Token {api_key} does not exist")
else:
print(isinstance(token.user_id, uuid.UUID), self.get_user(user_id=token.user_id))
return self.get_user(user_id=token.user_id)
@enforce_types
@@ -612,7 +605,6 @@ class MetadataStore:
)
for k, v in available_functions.items()
]
# print(results)
return results
# results = session.query(PresetModel).filter(PresetModel.user_id == user_id).all()
# return [r.to_record() for r in results]

5
memgpt/presets/presets.py
View File

@@ -41,11 +41,6 @@ def add_default_tools(user_id: uuid.UUID, ms: MetadataStore):
err = f"Error loading function set '{module_name}': {e}"
printd(err)
from pprint import pprint
print("BASE FUNCTIONS", functions_to_schema.keys())
pprint(functions_to_schema)
# create tool in db
for name, schema in functions_to_schema.items():
ms.add_tool(ToolModel(name=name, tags=["base"], source_type="python", json_schema=schema["json_schema"]))

101
memgpt/server/rest_api/admin/tools.py Normal file
View File

@@ -0,0 +1,101 @@
from typing import List, Literal, Optional
from fastapi import APIRouter, Body, HTTPException
from pydantic import BaseModel, Field
from memgpt.models.pydantic_models import ToolModel
from memgpt.server.rest_api.interface import QueuingInterface
from memgpt.server.server import SyncServer
router = APIRouter()
class ListToolsResponse(BaseModel):
tools: List[ToolModel] = Field(..., description="List of tools (functions).")
class CreateToolRequest(BaseModel):
name: str = Field(..., description="The name of the function.")
source_code: str = Field(..., description="The source code of the function.")
source_type: Optional[Literal["python"]] = Field(None, description="The type of the source code.")
tags: Optional[List[str]] = Field(None, description="Metadata tags.")
class CreateToolResponse(BaseModel):
tool: ToolModel = Field(..., description="Information about the newly created tool.")
def setup_tools_index_router(server: SyncServer, interface: QueuingInterface):
# get_current_user_with_server = partial(partial(get_current_user, server), password)
@router.delete("/tools/{tool_name}", tags=["tools"])
async def delete_tool(
tool_name: str,
# user_id: uuid.UUID = Depends(get_current_user_with_server), # TODO: add back when user-specific
):
"""
Delete a tool by name
"""
# Clear the interface
interface.clear()
# tool = server.ms.delete_tool(user_id=user_id, tool_name=tool_name) TODO: add back when user-specific
server.ms.delete_tool(name=tool_name)
@router.get("/tools/{tool_name}", tags=["tools"], response_model=ToolModel)
async def get_tool(tool_name: str):
"""
Get a tool by name
"""
# Clear the interface
interface.clear()
# tool = server.ms.get_tool(user_id=user_id, tool_name=tool_name) TODO: add back when user-specific
tool = server.ms.get_tool(tool_name=tool_name)
if tool is None:
# return 404 error
raise HTTPException(status_code=404, detail=f"Tool with name {tool_name} not found.")
return tool
@router.get("/tools", tags=["tools"], response_model=ListToolsResponse)
async def list_all_tools(
# user_id: uuid.UUID = Depends(get_current_user_with_server), # TODO: add back when user-specific
):
"""
Get a list of all tools available to agents created by a user
"""
# Clear the interface
interface.clear()
# tools = server.ms.list_tools(user_id=user_id) TODO: add back when user-specific
tools = server.ms.list_tools()
return ListToolsResponse(tools=tools)
@router.post("/tools", tags=["tools"], response_model=ToolModel)
async def create_tool(
request: CreateToolRequest = Body(...),
# user_id: uuid.UUID = Depends(get_current_user_with_server), # TODO: add back when user-specific
):
"""
Create a new tool (dummy route)
"""
from memgpt.functions.functions import load_function_file, write_function
# check if function already exists
if server.ms.get_tool(request.name):
raise ValueError(f"Tool with name {request.name} already exists.")
# write function to ~/.memgt/functions directory
file_path = write_function(request.name, request.name, request.source_code)
# TODO: Use load_function_file to load function schema
schema = load_function_file(file_path)
assert len(list(schema.keys())) == 1, "Function schema must have exactly one key"
json_schema = list(schema.values())[0]["json_schema"]
print("adding tool", request.name, request.tags, request.source_code)
tool = ToolModel(name=request.name, json_schema=json_schema, tags=request.tags, source_code=request.source_code)
tool.id
server.ms.add_tool(tool)
# TODO: insert tool information into DB as ToolModel
return server.ms.get_tool(request.name)
return router

1
memgpt/server/rest_api/agents/memory.py
View File

@@ -42,6 +42,7 @@ class ArchivalMemoryObject(BaseModel):
class GetAgentArchivalMemoryResponse(BaseModel):
# TODO: make this List[Passage] instead
archival_memory: List[ArchivalMemoryObject] = Field(..., description="A list of all memory objects in archival memory.")

8
memgpt/server/rest_api/server.py
View File

@@ -12,6 +12,7 @@ from fastapi.security import HTTPAuthorizationCredentials, HTTPBearer
from starlette.middleware.cors import CORSMiddleware
from memgpt.server.constants import REST_DEFAULT_PORT
from memgpt.server.rest_api.admin.tools import setup_tools_index_router
from memgpt.server.rest_api.admin.users import setup_admin_router
from memgpt.server.rest_api.agents.command import setup_agents_command_router
from memgpt.server.rest_api.agents.config import setup_agents_config_router
@@ -30,7 +31,7 @@ from memgpt.server.rest_api.personas.index import setup_personas_index_router
from memgpt.server.rest_api.presets.index import setup_presets_index_router
from memgpt.server.rest_api.sources.index import setup_sources_index_router
from memgpt.server.rest_api.static_files import mount_static_files
from memgpt.server.rest_api.tools.index import setup_tools_index_router
from memgpt.server.rest_api.tools.index import setup_user_tools_index_router
from memgpt.server.server import SyncServer
from memgpt.settings import settings
@@ -81,6 +82,7 @@ app.include_router(setup_auth_router(server, interface, password), prefix=API_PR
# /admin/users endpoints
app.include_router(setup_admin_router(server, interface), prefix=ADMIN_PREFIX, dependencies=[Depends(verify_password)])
app.include_router(setup_tools_index_router(server, interface), prefix=ADMIN_PREFIX, dependencies=[Depends(verify_password)])
# /api/agents endpoints
app.include_router(setup_agents_command_router(server, interface, password), prefix=API_PREFIX)
@@ -91,9 +93,7 @@ app.include_router(setup_agents_message_router(server, interface, password), pre
app.include_router(setup_humans_index_router(server, interface, password), prefix=API_PREFIX)
app.include_router(setup_personas_index_router(server, interface, password), prefix=API_PREFIX)
app.include_router(setup_models_index_router(server, interface, password), prefix=API_PREFIX)
app.include_router(
setup_tools_index_router(server, interface, password), prefix=API_PREFIX, dependencies=[Depends(verify_password)]
) # admin only
app.include_router(setup_user_tools_index_router(server, interface, password), prefix=API_PREFIX)
app.include_router(setup_sources_index_router(server, interface, password), prefix=API_PREFIX)
app.include_router(setup_presets_index_router(server, interface, password), prefix=API_PREFIX)

59
memgpt/server/rest_api/tools/index.py
View File

@@ -1,9 +1,12 @@
import uuid
from functools import partial
from typing import List, Literal, Optional
from fastapi import APIRouter, Body, HTTPException
from fastapi import APIRouter, Depends, HTTPException
from pydantic import BaseModel, Field
from memgpt.models.pydantic_models import ToolModel
from memgpt.server.rest_api.auth_token import get_current_user
from memgpt.server.rest_api.interface import QueuingInterface
from memgpt.server.server import SyncServer
@@ -25,24 +28,14 @@ class CreateToolResponse(BaseModel):
tool: ToolModel = Field(..., description="Information about the newly created tool.")
def setup_tools_index_router(server: SyncServer, interface: QueuingInterface, password: str):
# get_current_user_with_server = partial(partial(get_current_user, server), password)
@router.delete("/tools/{tool_name}", tags=["tools"])
async def delete_tool(
tool_name: str,
# user_id: uuid.UUID = Depends(get_current_user_with_server), # TODO: add back when user-specific
):
"""
Delete a tool by name
"""
# Clear the interface
interface.clear()
# tool = server.ms.delete_tool(user_id=user_id, tool_name=tool_name) TODO: add back when user-specific
server.ms.delete_tool(name=tool_name)
def setup_user_tools_index_router(server: SyncServer, interface: QueuingInterface, password: str):
get_current_user_with_server = partial(partial(get_current_user, server), password)
@router.get("/tools/{tool_name}", tags=["tools"], response_model=ToolModel)
async def get_tool(tool_name: str):
async def get_tool(
tool_name: str,
user_id: uuid.UUID = Depends(get_current_user_with_server), # TODO: add back when user-specific
):
"""
Get a tool by name
"""
@@ -57,7 +50,7 @@ def setup_tools_index_router(server: SyncServer, interface: QueuingInterface, pa
@router.get("/tools", tags=["tools"], response_model=ListToolsResponse)
async def list_all_tools(
# user_id: uuid.UUID = Depends(get_current_user_with_server), # TODO: add back when user-specific
user_id: uuid.UUID = Depends(get_current_user_with_server), # TODO: add back when user-specific
):
"""
Get a list of all tools available to agents created by a user
@@ -68,34 +61,4 @@ def setup_tools_index_router(server: SyncServer, interface: QueuingInterface, pa
tools = server.ms.list_tools()
return ListToolsResponse(tools=tools)
@router.post("/tools", tags=["tools"], response_model=ToolModel)
async def create_tool(
request: CreateToolRequest = Body(...),
# user_id: uuid.UUID = Depends(get_current_user_with_server), # TODO: add back when user-specific
):
"""
Create a new tool (dummy route)
"""
from memgpt.functions.functions import load_function_file, write_function
# check if function already exists
if server.ms.get_tool(request.name):
raise ValueError(f"Tool with name {request.name} already exists.")
# write function to ~/.memgt/functions directory
file_path = write_function(request.name, request.name, request.source_code)
# TODO: Use load_function_file to load function schema
schema = load_function_file(file_path)
assert len(list(schema.keys())) == 1, "Function schema must have exactly one key"
json_schema = list(schema.values())[0]["json_schema"]
print("adding tool", request.name, request.tags, request.source_code)
tool = ToolModel(name=request.name, json_schema=json_schema, tags=request.tags, source_code=request.source_code)
tool.id
server.ms.add_tool(tool)
# TODO: insert tool information into DB as ToolModel
return server.ms.get_tool(request.name)
return router

3
memgpt/server/server.py
View File

@@ -137,7 +137,6 @@ class LockingServer(Server):
try:
# Execute the function
# logger.info(f"running function on agent_id = {agent_id}")
print("USERID", user_id)
return func(self, user_id, agent_id, *args, **kwargs)
finally:
# Release the lock
@@ -1054,7 +1053,6 @@ class SyncServer(LockingServer):
# Assume passages
records = memgpt_agent.persistence_manager.archival_memory.storage.get_all()
print("records:", records)
return [dict(id=str(r.id), contents=r.text) for r in records]
@@ -1269,7 +1267,6 @@ class SyncServer(LockingServer):
def api_key_to_user(self, api_key: str) -> uuid.UUID:
"""Decode an API key to a user"""
user = self.ms.get_user_from_api_key(api_key=api_key)
print("got user", api_key, user.id)
if user is None:
raise HTTPException(status_code=403, detail="Invalid credentials")
else:

1
pyproject.toml
View File

@@ -79,6 +79,7 @@ ollama = ["llama-index-embeddings-ollama"]
[tool.black]
line-length = 140
target-version = ['py310', 'py311']
extend-exclude = "examples/*"
[build-system]
requires = ["poetry-core"]

1
tests/test_client.py
View File

@@ -170,6 +170,7 @@ def test_archival_memory(client, agent):
assert insert_response, "Inserting archival memory failed"
archival_memory_response = client.get_agent_archival_memory(agent_id=agent.id, limit=1)
print("MEMORY")
archival_memories = [memory.contents for memory in archival_memory_response.archival_memory]
assert memory_content in archival_memories, f"Retrieving archival memory failed: {archival_memories}"

8
tests/test_summarize.py
View File

@@ -47,25 +47,25 @@ def test_summarize():
response = client.user_message(
agent_id=agent_obj.agent_state.id,
message="Hey, how's it going? What do you think about this whole shindig",
)
).messages
assert response is not None and len(response) > 0
print(f"test_summarize: response={response}")
response = client.user_message(
agent_id=agent_obj.agent_state.id,
message="Any thoughts on the meaning of life?",
)
).messages
assert response is not None and len(response) > 0
print(f"test_summarize: response={response}")
response = client.user_message(agent_id=agent_obj.agent_state.id, message="Does the number 42 ring a bell?")
response = client.user_message(agent_id=agent_obj.agent_state.id, message="Does the number 42 ring a bell?").messages
assert response is not None and len(response) > 0
print(f"test_summarize: response={response}")
response = client.user_message(
agent_id=agent_obj.agent_state.id,
message="Would you be surprised to learn that you're actually conversing with an AI right now?",
)
).messages
assert response is not None and len(response) > 0
print(f"test_summarize: response={response}")