Skip to content

Agent Zero AI framework

memory tool update, minor cleanup, utility models

frdelb3adcc7bca8e
diff --git a/agent.py b/agent.py
index 1563ecfed..a2ff200d0 100644
--- a/agent.py
+++ b/agent.py
@@ -1,7 +1,6 @@
from dataclasses import dataclass, field
import time, importlib, inspect, os, json
-import traceback
-from typing import Any, Optional, Dict, TypedDict
+from typing import Any, Optional, Dict
from python.helpers import extract_tools, rate_limiter, files, errors
from python.helpers.print_style import PrintStyle
from langchain.schema import AIMessage
@@ -9,20 +8,20 @@ from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_core.messages import HumanMessage, SystemMessage
from langchain_core.language_models.chat_models import BaseChatModel
from langchain_core.embeddings import Embeddings
-from python.helpers.rate_limiter import RateLimiter
# rate_limit = rate_limiter.rate_limiter(30,160000) #TODO! implement properly
@dataclass
class AgentConfig:
chat_model:BaseChatModel
+ utility_model: BaseChatModel
embeddings_model:Embeddings
memory_subdir: str = ""
auto_memory_count: int = 3
auto_memory_skip: int = 2
rate_limit_seconds: int = 60
- rate_limit_requests: int = 30
- rate_limit_input_tokens: int = 0
+ rate_limit_requests: int = 15
+ rate_limit_input_tokens: int = 1000000
rate_limit_output_tokens: int = 0
msgs_keep_max: int = 25
msgs_keep_start: int = 5
@@ -62,7 +61,7 @@ class Agent:
self.last_message = ""
self.intervention_message = ""
self.intervention_status = False
- self.rate_limiter = RateLimiter(max_calls=self.config.rate_limit_requests,max_input_tokens=self.config.rate_limit_input_tokens,max_output_tokens=self.config.rate_limit_output_tokens,window_seconds=self.config.rate_limit_seconds)
+ self.rate_limiter = rate_limiter.RateLimiter(max_calls=self.config.rate_limit_requests,max_input_tokens=self.config.rate_limit_input_tokens,max_output_tokens=self.config.rate_limit_output_tokens,window_seconds=self.config.rate_limit_seconds)
self.data = {} # free data object all the tools can use
os.chdir(files.get_abs_path("./work_dir")) #change CWD to work_dir
@@ -160,7 +159,7 @@ class Agent:
SystemMessage(content=system),
HumanMessage(content=msg)])
- chain = prompt | self.config.chat_model
+ chain = prompt | self.config.utility_model
response = ""
printer = None
diff --git a/example.env b/example.env
index 2aa725749..342a8765b 100644
--- a/example.env
+++ b/example.env
@@ -2,7 +2,7 @@ API_KEY_OPENAI=
API_KEY_ANTHROPIC=
API_KEY_GROQ=
API_KEY_PERPLEXITY=
-
+API_KEY_GOOGLE=
TOKENIZERS_PARALLELISM=true
PYDEVD_DISABLE_FILE_VALIDATION=1
\ No newline at end of file
diff --git a/main.py b/main.py
index d7527441d..a57269806 100644
--- a/main.py
+++ b/main.py
@@ -1,8 +1,5 @@
import signal
-import sys
-import termios
import threading, time, models, os
-import tty
from ansio import application_keypad, mouse_input, raw_input
from ansio.input import InputEvent, get_input_event
from agent import Agent, AgentConfig
@@ -33,6 +30,9 @@ def initialize():
# chat_llm = models.get_ollama(model_name="qwen:14b")
chat_llm = models.get_google_chat()
+ utility_llm = models.get_anthropic_haiku(temperature=0)
+
+
# embedding model used for memory
# embedding_llm = models.get_embedding_openai()
embedding_llm = models.get_embedding_hf()
@@ -40,6 +40,7 @@ def initialize():
# agent configuration
config = AgentConfig(
chat_model = chat_llm,
+ utility_model = utility_llm,
embeddings_model = embedding_llm,
# memory_subdir = "",
auto_memory_count = 0,
@@ -81,13 +82,13 @@ def chat(agent:Agent):
with input_lock:
timeout = agent.get_data("timeout") # how long the agent is willing to wait
if not timeout: # if agent wants to wait for user input forever
- PrintStyle(background_color="#6C3483", font_color="white", bold=True, padding=True).print(f"User message ('exit' to leave):")
+ PrintStyle(background_color="#6C3483", font_color="white", bold=True, padding=True).print(f"User message ('e' to leave):")
import readline # this fixes arrow keys in terminal
user_input = input("> ")
PrintStyle(font_color="white", padding=False, log_only=True).print(f"> {user_input}")
else: # otherwise wait for user input with a timeout
- PrintStyle(background_color="#6C3483", font_color="white", bold=True, padding=True).print(f"User message ({timeout}s timeout, 'wait' to wait, 'exit' to leave):")
+ PrintStyle(background_color="#6C3483", font_color="white", bold=True, padding=True).print(f"User message ({timeout}s timeout, 'w' to wait, 'e' to leave):")
import readline # this fixes arrow keys in terminal
# user_input = timed_input("> ", timeout=timeout)
user_input = timeout_input("> ", timeout=timeout)
@@ -97,14 +98,14 @@ def chat(agent:Agent):
PrintStyle(font_color="white", padding=False).stream(f"{user_input}")
else:
user_input = user_input.strip()
- if user_input.lower()=="wait": # the user needs more time
+ if user_input.lower()=="w": # the user needs more time
user_input = input("> ").strip()
PrintStyle(font_color="white", padding=False, log_only=True).print(f"> {user_input}")
# exit the conversation when the user types 'exit'
- if user_input.lower() == 'exit': break
+ if user_input.lower() == 'e': break
# send message to agent0,
assistant_response = agent.message_loop(user_input)
@@ -118,13 +119,13 @@ def chat(agent:Agent):
def intervention():
if Agent.streaming_agent and not Agent.paused:
Agent.paused = True # stop agent streaming
- PrintStyle(background_color="#6C3483", font_color="white", bold=True, padding=True).print(f"User intervention ('exit' to leave, empty to continue):")
+ PrintStyle(background_color="#6C3483", font_color="white", bold=True, padding=True).print(f"User intervention ('e' to leave, empty to continue):")
import readline # this fixes arrow keys in terminal
user_input = input("> ").strip()
PrintStyle(font_color="white", padding=False, log_only=True).print(f"> {user_input}")
- if user_input.lower() == 'exit': os._exit(0) # exit the conversation when the user types 'exit'
+ if user_input.lower() == 'e': os._exit(0) # exit the conversation when the user types 'exit'
if user_input: Agent.streaming_agent.intervention_message = user_input # set intervention message if non-empty
Agent.paused = False # continue agent streaming
diff --git a/models.py b/models.py
index 321041c3a..22fbe4b09 100755
--- a/models.py
+++ b/models.py
@@ -5,7 +5,7 @@ from langchain_openai import ChatOpenAI, OpenAI, OpenAIEmbeddings
from langchain_anthropic import ChatAnthropic
from langchain_groq import ChatGroq
from langchain_huggingface import HuggingFaceEmbeddings
-from langchain_google_genai import ChatGoogleGenerativeAI
+from langchain_google_genai import ChatGoogleGenerativeAI, HarmBlockThreshold, HarmCategory
# Load environment variables
@@ -76,15 +76,16 @@ def get_groq_gemma(api_key=None, temperature=DEFAULT_TEMPERATURE):
api_key = api_key or get_api_key("groq")
return ChatGroq(model_name="gemma-7b-it", temperature=temperature, api_key=api_key) # type: ignore
-def get_ollama_dolphin(api_key=None, temperature=DEFAULT_TEMPERATURE):
+def get_ollama_dolphin(temperature=DEFAULT_TEMPERATURE):
return Ollama(model="dolphin-llama3:8b-256k-v2.9-fp16", temperature=temperature)
-def get_ollama_phi(api_key=None, temperature=DEFAULT_TEMPERATURE):
+def get_ollama_phi(temperature=DEFAULT_TEMPERATURE):
return Ollama(model="phi3:3.8b-mini-instruct-4k-fp16",temperature=temperature)
def get_google_chat(model_name="gemini-1.5-flash-latest", api_key=None, temperature=DEFAULT_TEMPERATURE):
api_key = api_key or get_api_key("google")
- return ChatGoogleGenerativeAI(model=model_name, temperature=temperature, google_api_key=api_key) # type: ignore
+ return ChatGoogleGenerativeAI(model=model_name, temperature=temperature, google_api_key=api_key,
+ safety_settings={HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.BLOCK_NONE }) # type: ignore
def get_embedding_hf(model_name="sentence-transformers/all-MiniLM-L6-v2"):
return HuggingFaceEmbeddings(model_name=model_name)
diff --git a/prompts/agent.tools.md b/prompts/agent.tools.md
index 3c4e0101b..f6e2042f9 100644
--- a/prompts/agent.tools.md
+++ b/prompts/agent.tools.md
@@ -62,11 +62,28 @@ Always verify memory by online.
}
~~~
-### memorize:
-Save information to persistent memory.
+### memory_tool:
+Manage long term memories. Allowed arguments are "query", "memorize" and "forget".
Memories can help you remember important details and later reuse them.
+When querying, provide a "query" argument to search for. You will retrieve IDs and contents of relevant memories. Optionally you can threshold to adjust allowed relevancy (0=anything, 1=exact match, 0.1 is default).
+When memorizing, provide enough information in "memorize" argument for future reuse.
+When forgetting, provide memory IDs from loaded memories separated by commas in "forget" argument.
Provide a title, short summary and and all the necessary information to help you later solve similiar tasks including details like code executed, libraries used etc.
**Example usages**:
+1. load:
+~~~json
+{
+ "thoughts": [
+ "Let's search my memory for...",
+ ],
+ "tool_name": "memory_tool",
+ "tool_args": {
+ "query": "File compression library for...",
+ "threshold": 0.1
+ }
+}
+~~~
+2. save:
~~~json
{
"thoughts": [
@@ -74,9 +91,21 @@ Provide a title, short summary and and all the necessary information to help you
"Details of this process will be valuable...",
"Let's save tools and code used...",
],
- "tool_name": "memorize",
+ "tool_name": "memory_tool",
"tool_args": {
- "memory": "# How to...",
+ "memorize": "# How to...",
+ }
+}
+~~~
+3. delete:
+~~~json
+{
+ "thoughts": [
+ "User asked to delete memories...",
+ ],
+ "tool_name": "memory_tool",
+ "tool_args": {
+ "forget": "32cd37ffd1-101f-4112-80e2-33b795548116, d1306e36-6a9c-4e6a-bfc3-c8335035dcf8 ...",
}
}
~~~
@@ -93,7 +122,8 @@ When tool outputs error, you need to change your code accordingly before trying
Keep in mind that current working directory CWD automatically resets before every tool call.
IMPORTANT!: Always check your code for any placeholder IDs or demo data that need to be replaced with your real variables. Do not simply reuse code snippets from tutorials.
Do not use in combination with other tools except for thoughts. Wait for response before using other tools.
-**Example usage**:
+**Example usages:**
+1. Execute python code
~~~json
{
"thoughts": [
@@ -101,10 +131,52 @@ Do not use in combination with other tools except for thoughts. Wait for respons
"I can use library...",
"Then I can...",
],
- "tool_name": "memory_tool",
+ "tool_name": "code_execution_tool",
"tool_args": {
"runtime": "python",
"code": "import os\nreturn os.getcwd()",
}
}
+~~~
+
+2. Execute terminal command
+~~~json
+{
+ "thoughts": [
+ "I need to do...",
+ "I need to install...",
+ ],
+ "tool_name": "code_execution_tool",
+ "tool_args": {
+ "runtime": "terminal",
+ "code": "apt-get install zip",
+ }
+}
+~~~
+
+2. 1. Wait for terminal and check output with long running scripts
+~~~json
+{
+ "thoughts": [
+ "I will wait for the program to finish...",
+ ],
+ "tool_name": "code_execution_tool",
+ "tool_args": {
+ "runtime": "output",
+ }
+}
+~~~
+
+2. 2. Answer terminal dialog
+~~~json
+{
+ "thoughts": [
+ "Program needs confirmation...",
+ ],
+ "tool_name": "code_execution_tool",
+ "tool_args": {
+ "runtime": "terminal",
+ "code": "Y",
+ }
+}
~~~
\ No newline at end of file
diff --git a/python/helpers/vdb.py b/python/helpers/vdb.py
new file mode 100644
index 000000000..325112091
--- /dev/null
+++ b/python/helpers/vdb.py
@@ -0,0 +1,69 @@
+from langchain.storage import InMemoryByteStore, LocalFileStore
+from langchain.embeddings import CacheBackedEmbeddings
+from langchain_core.embeddings import Embeddings
+
+from langchain_chroma import Chroma
+import chromadb
+from chromadb.config import Settings
+
+from . import files
+from langchain_core.documents import Document
+import uuid
+
+
+class VectorDB:
+
+ def __init__(self, embeddings_model:Embeddings, in_memory=False, cache_dir="./cache"):
+ print("Initializing VectorDB...")
+ self.embeddings_model = embeddings_model
+
+ db_cache = files.get_abs_path(cache_dir,"database")
+
+ self.client =chromadb.PersistentClient(path=db_cache)
+ self.collection = self.client.create_collection("my_collection")
+ self.collection
+
+
+ def search(self, query:str, results=2):
+ emb = self.embeddings_model.embed_query(query)
+ res = self.collection.query(query_embeddings=[emb],n_results=results)
+ best = res["documents"][0][0] # type: ignore
+
+ # def delete_documents(self, query):
+ # score_limit = 1
+ # k = 2
+ # tot = 0
+ # while True:
+ # # Perform similarity search with score
+ # docs = self.db.similarity_search_with_score(query, k=k)
+
+ # # Extract document IDs and filter based on score
+ # document_ids = [result[0].metadata["id"] for result in docs if result[1] < score_limit]
+
+ # # Delete documents with IDs over the threshold score
+ # if document_ids:
+ # fnd = self.db.get(where={"id": {"$in": document_ids}})
+ # if fnd["ids"]: self.db.delete(ids=fnd["ids"])
+ # tot += len(fnd["ids"])
+
+ # # If fewer than K document IDs, break the loop
+ # if len(document_ids) < k:
+ # break
+
+ # return tot
+
+ def insert(self, data:str):
+
+ id = str(uuid.uuid4())
+ emb = self.embeddings_model.embed_documents([data])[0]
+
+ self.collection.add(
+ ids=[id],
+ embeddings=[emb],
+ documents=[data],
+ )
+
+ return id
+
+
+
diff --git a/python/helpers/vector_db.py b/python/helpers/vector_db.py
index 2bcc6ae91..4ffdb376b 100644
--- a/python/helpers/vector_db.py
+++ b/python/helpers/vector_db.py
@@ -1,6 +1,7 @@
from langchain.storage import InMemoryByteStore, LocalFileStore
from langchain.embeddings import CacheBackedEmbeddings
from langchain_chroma import Chroma
+
from . import files
from langchain_core.documents import Document
import uuid
@@ -29,8 +30,12 @@ class VectorDB:
self.db = Chroma(embedding_function=self.embedder,persist_directory=db_cache)
+
def search_similarity(self, query, results=3):
return self.db.similarity_search(query,results)
+
+ def search_similarity_threshold(self, query, results=3, threshold=0.5):
+ return self.db.search(query,search_type="similarity_score_threshold",score_threshold=threshold)
def search_max_rel(self, query, results=3):
return self.db.max_marginal_relevance_search(query,results)
@@ -60,7 +65,8 @@ class VectorDB:
def insert_document(self, data):
id = str(uuid.uuid4())
- self.db.add_documents(documents=[ Document(data, metadata={"id": id}) ])
+ self.db.add_documents(documents=[ Document(data, metadata={"id": id}) ], ids=[id])
+
return id
diff --git a/python/tools/code_execution_tool.py b/python/tools/code_execution_tool.py
index c64dc6bd3..d9c2cdc80 100644
--- a/python/tools/code_execution_tool.py
+++ b/python/tools/code_execution_tool.py
@@ -32,6 +32,8 @@ class CodeExecution(Tool):
response = self.execute_nodejs_code(self.args["code"])
elif runtime == "terminal":
response = self.execute_terminal_command(self.args["code"])
+ elif runtime == "output":
+ response = self.get_terminal_output()
else:
response = files.read_file("./prompts/fw.code_runtime_wrong.md", runtime=runtime)
@@ -78,7 +80,9 @@ class CodeExecution(Tool):
self.state.shell.send_command(command)
PrintStyle(background_color="white",font_color="#85C1E9",bold=True).print(f"{self.agent.agent_name} code execution output:")
+ return self.get_terminal_output()
+ def get_terminal_output(self):
idle=0
while True:
time.sleep(0.1) # Wait for some output to be generated
diff --git a/python/tools/memorize.py b/python/tools/memorize.py
deleted file mode 100644
index bdef2032c..000000000
--- a/python/tools/memorize.py
+++ /dev/null
@@ -1,14 +0,0 @@
-from agent import Agent
-from python.helpers import files
-from python.helpers.tool import Tool, Response
-import memory_tool
-
-class Memorize(Tool):
- def execute(self,**kwargs):
-
- memory_tool.process_query(self.agent, str(self.args), "save")
-
- return Response(
- message=files.read_file("prompts/fw.memorized.md"),
- break_loop=False,
- )
\ No newline at end of file
diff --git a/python/tools/memory_tool.py b/python/tools/memory_tool.py
index 31eb1ec7f..2abdf81d1 100644
--- a/python/tools/memory_tool.py
+++ b/python/tools/memory_tool.py
@@ -9,7 +9,18 @@ db: VectorDB | None = None
class Memory(Tool):
def execute(self,**kwargs):
- result = process_query(self.agent, self.args["memory"],self.args["action"], result_count=self.agent.config.auto_memory_count)
+ result=[]
+
+ if "query" in kwargs:
+ if "threshold" in kwargs: threshold = float(kwargs["threshold"])
+ else: threshold = 0.1
+ result = process_query(self.agent, kwargs["query"], action="load", threshold=threshold, result_count=3)
+ elif "memorize" in kwargs:
+ result = process_query(self.agent, kwargs["memorize"], action="save")
+ elif "forget" in kwargs:
+ result = process_query(self.agent, kwargs["forget"], action="delete")
+
+ # result = process_query(self.agent, self.args["memory"],self.args["action"], result_count=self.agent.config.auto_memory_count)
return Response(message="\n\n".join(result), break_loop=False)
@@ -19,21 +30,21 @@ def initialize(embeddings_model, subdir=""):
db = VectorDB(embeddings_model=embeddings_model, in_memory=False, cache_dir=dir)
-def process_query(agent:Agent, message: str, action: str = "load", result_count: int = 3, **kwargs):
+def process_query(agent:Agent, message: str, action: str = "load", result_count: int = 3, threshold: float = 0.1, **kwargs):
if not db: initialize(agent.config.embeddings_model, subdir=agent.config.memory_subdir)
if action.strip().lower() == "save":
id = db.insert_document(str(message)) # type: ignore
- return files.read_file("./prompts/fw.memory_saved.md")
+ return [files.read_file("./prompts/fw.memory_saved.md")]
elif action.strip().lower() == "delete":
deleted = db.delete_documents(message) # type: ignore
- return files.read_file("./prompts/fw.memories_deleted.md", count=deleted)
+ return [files.read_file("./prompts/fw.memories_deleted.md", count=deleted)]
else:
results=[]
- docs = db.search_max_rel(message,result_count) # type: ignore
- if len(docs)==0: return files.read_file("./prompts/fw.memories_not_found.md", query=message)
+ docs = db.search_similarity_threshold(message,result_count,threshold) # type: ignore
+ if len(docs)==0: return [files.read_file("./prompts/fw.memories_not_found.md", query=message)]
for doc in docs:
results.append(doc.page_content)
return results
diff --git a/requirements.txt b/requirements.txt
index f27283000..44fced019 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,14 +1,14 @@
ansio==0.0.1
python-dotenv==1.0.1
-langchain-groq==0.1.5
+langchain-groq==0.1.6
langchain-huggingface==0.0.3
-langchain-openai==0.1.8
-langchain-community==0.2.4
-langchain-anthropic==0.1.15
-langchain-chroma==0.1.1
-google-generativeai==0.5.4
+langchain-openai==0.1.15
+langchain-community==0.2.7
+langchain-anthropic==0.1.19
+langchain-chroma==0.1.2
+langchain-google-genai==1.0.7
webcolors==24.6.0
sentence-transformers==3.0.1
pytimedinput==2.0.1
docker==7.1.0
-paramiko==3.4.0
+paramiko==3.4.0
\ No newline at end of file