llm/SIA
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge branch 'master' of https://git.nielsgeens.be/llm/SIA

Browse Source
This commit is contained in:
Niels Geens
2024-11-13 11:55:06 +01:00
parent 555fc7f764 892f2b7bc7
commit d0eed2666e
7 changed files with 127 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
action_schema.xsd
View File

@@ -1,4 +1,7 @@
<?xml version="1.0" encoding="UTF-8"?>
<!--
Always answer with a single element.
-->
<xs:schema xmlns:xs="http://www.w3.org/2001/XMLSchema" elementFormDefault="qualified">
<!--
@@ -6,6 +9,8 @@
Use it to remove unnecessary items and stop background processes.
When you delete something, it is gone.
Make sure all important info is stored in files.
Example:
<delete id="1234567890"/>
-->
<xs:element name="delete">
<xs:complexType>
@@ -17,6 +22,8 @@
Stop command terminates the agent gracefully.
For the main SIA instance this will trigger an update and restart.
For sub-instances this is the correct way to stop after all tasks are complete.
Example:
<stop id="1234567890"/>
-->
<xs:element name="stop">
<xs:complexType/>
@@ -26,6 +33,10 @@
Single script that runs once and completes.
Output is stored in context until explicitly deleted.
Used for one-time operations like file manipulation.
Example:
<single>
ls /
</single>
-->
<xs:element name="single">
<xs:complexType mixed="true">
@@ -40,6 +51,10 @@
After a command is issued, all repeat scripts in context are run again.
Useful for monitoring changing files or viewing results immediately after changing a file.
Repeat scripts should execute quickly to avoid blocking the agent.
Example:
<repeat>
ls /
</repeat>
-->
<xs:element name="repeat">
<xs:complexType mixed="true">
@@ -53,6 +68,10 @@
As an agent it is important to reason about your actions and their results.
In a reasoning action you can write freeform text.
This is also stored in context until deleted.
Example:
<reasoning>
I should explore the file system for interesting files.
</reasoning>
-->
<xs:element name="reasoning">
<xs:complexType mixed="true">
@@ -64,6 +83,8 @@
<!--
Read all available text on stdin and store it in context.
Example:
<read_stdin/>
-->
<xs:element name="read_stdin">
<xs:complexType/>
@@ -72,6 +93,10 @@
<!--
Write to stdout.
This is your main way of contacting the user.
Example:
<write_stdout>
Hello world!
</write_stdout>
-->
<xs:element name="write_stdout">
<xs:complexType mixed="true">

1
requirements.txt
View File

@@ -1,6 +1,7 @@
accelerate
aiohttp
bs4
openai
python-dotenv
torch
transformers

7
sia/__main__.py
View File

@@ -5,11 +5,11 @@ import asyncio
import mimetypes
import time
from .config import Config
from .hf_llm_engine import HfLlmEngine
from .llm_engine import LlmEngine
from .local_llm_engine import LocalLlmEngine
from .openai_llm_engine import OpenAILlmEngine
from .response_parser import ResponseParser
from .system_metrics import SystemMetrics
from .web_agent import WebAgent
@@ -51,6 +51,11 @@ class Main:
api_token=self._config.api_token,
temperature=self._config.temperature
)
case "openai":
self._llm = OpenAILlmEngine(
model=self._config.model,
api_key=self._config.api_token
)
case "test":
self._llm = TestLLM()
case _:

57
sia/openai_llm_engine.py Normal file
View File

@@ -0,0 +1,57 @@
from typing import Iterator, Optional
import openai
import json
from .llm_engine import LlmEngine
class OpenAILlmEngine(LlmEngine):
"""
LLM Engine implementation using OpenAI's API.
Supports streaming responses from chat completion models.
"""
def __init__(
self,
model: str,
api_key: str,
):
"""
Initialize the OpenAI LLM Engine.
Args:
model: OpenAI model to use (default: gpt-4)
api_key: OpenAI API key. If None, will try to read from OPENAI_API_KEY env var
"""
self._model = model
# Initialize OpenAI client
self.client = openai.Client(
api_key=api_key,
)
def infer(self, system_prompt: str, main_context: str) -> Iterator[str]:
"""
Run inference using the system prompt and main context.
Args:
system_prompt: The system prompt string
main_context: The main context string after templating
Returns:
Iterator[str]: An iterator that yields the generated text in chunks.
"""
messages = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": main_context}
]
stream = self.client.chat.completions.create(
model=self._model,
messages=messages,
stream=True,
temperature=0.3,
)
for chunk in stream:
if content := chunk.choices[0].delta.content:
yield content

96
sia/response_parser.py
View File

@@ -37,46 +37,6 @@ class ResponseParser:
"""
self.io_buffer = io_buffer
def _extract_content(self, xml: str, root: ET.Element) -> str:
"""
Extract content from XML, handling CDATA sections and preserving special characters.
Args:
xml: Original XML string
root: Parsed XML element
Returns:
str: Extracted content with proper handling of CDATA and special characters
"""
if root.text is None:
return ""
# Find opening tag end
tag_name = root.tag
tag_start = xml.find('<' + tag_name)
tag_end = xml.find('>', tag_start)
if tag_end == -1:
return ""
# Check for self-closing tag
if xml[tag_end-1] == '/':
return ""
# Find closing tag start
close_tag = '</' + tag_name + '>'
close_start = xml.rfind(close_tag)
if close_start == -1:
return ""
# Extract content
content = xml[tag_end+1:close_start]
# Handle CDATA sections
if content.startswith('<![CDATA[') and content.endswith(']]>'):
content = content[9:-3] # Remove CDATA markers
return content
def parse(self, xml: str) -> Union[Command, Entry]:
"""
Parse XML response into a Command or Entry.
@@ -91,74 +51,50 @@ class ResponseParser:
entry_id = str(uuid.uuid4())
timestamp = datetime.now()
# First try to parse the XML
try:
root = ET.fromstring(xml.strip())
except ET.ParseError as e:
return ParseErrorEntry(xml, f"Invalid XML: {str(e)}", entry_id, timestamp)
# Extract content preserving special characters and handling CDATA
content = self._extract_content(xml, root)
# Convert to appropriate type based on root tag
try:
if root.tag == 'delete':
# Extract target ID and create delete command
target_id = root.get('id')
if not target_id:
return ParseErrorEntry(xml, "Delete command missing required 'id' attribute", entry_id, timestamp)
if len(root.attrib) > 1:
return ParseErrorEntry(xml, "Delete command should only have 'id' attribute", entry_id, timestamp)
return DeleteCommand(target_id)
elif root.tag == 'stop':
# Create stop command
return StopCommand()
elif root.tag == 'background':
# Create background script entry
if not content:
return ParseErrorEntry(xml, "Background entry missing script content", entry_id, timestamp)
if root.attrib:
return ParseErrorEntry(xml, "Background entry shouldn't have attributes", entry_id, timestamp)
return BackgroundEntry(content, entry_id, timestamp)
if len(root) != 0 or root.attrib or root.text is None or root.text.strip() == '':
return ParseErrorEntry(xml, "Background entry requires (only) script content", entry_id, timestamp)
return BackgroundEntry(root.text, entry_id, timestamp)
elif root.tag == 'repeat':
# Create repeat script entry
if not content:
return ParseErrorEntry(xml, "Repeat entry missing script content", entry_id, timestamp)
if root.attrib:
return ParseErrorEntry(xml, "Repeat entry shouldn't have attributes", entry_id, timestamp)
return RepeatEntry(content, entry_id, timestamp)
if len(root) != 0 or root.attrib or root.text is None or root.text.strip() == '':
return ParseErrorEntry(xml, "Repeat entry requires (only) script content", entry_id, timestamp)
return RepeatEntry(root.text, entry_id, timestamp)
elif root.tag == 'single':
# Create single shot script entry
if not content:
return ParseErrorEntry(xml, "Single entry missing script content", entry_id, timestamp)
if root.attrib:
return ParseErrorEntry(xml, "Single entry shouldn't have attributes", entry_id, timestamp)
return SingleEntry(content, entry_id, timestamp)
if len(root) != 0 or root.attrib or root.text is None or root.text.strip() == '':
return ParseErrorEntry(xml, "Single entry requires (only) script content", entry_id, timestamp)
return SingleEntry(root.text, entry_id, timestamp)
elif root.tag == 'reasoning':
# Create reasoning entry
if not content:
return ParseErrorEntry(xml, "Reasoning entry missing content", entry_id, timestamp)
if root.attrib:
return ParseErrorEntry(xml, "Reasoning entry shouldn't have attributes", entry_id, timestamp)
return ReasoningEntry(content, entry_id, timestamp)
if len(root) != 0 or root.attrib or root.text is None or root.text.strip() == '':
return ParseErrorEntry(xml, "Reasoning entry requires (only) text content", entry_id, timestamp)
return ReasoningEntry(root.text, entry_id, timestamp)
elif root.tag == 'read_stdin':
# Create read entry - no content required
if root.attrib:
return ParseErrorEntry(xml, "Read stdin entry shouldn't have attributes", entry_id, timestamp)
return ReadEntry(self.io_buffer, entry_id, timestamp)
elif root.tag == 'write_stdout':
# Create write entry
if not content:
return ParseErrorEntry(xml, "Write entry missing content", entry_id, timestamp)
if root.attrib:
return ParseErrorEntry(xml, "Read stdin entry shouldn't have attributes", entry_id, timestamp)
return WriteEntry(content, self.io_buffer, entry_id, timestamp)
if len(root) != 0 or root.attrib or root.text is None or root.text.strip() == '':
return ParseErrorEntry(xml, "Write stdout entry requires (only) text content", entry_id, timestamp)
return WriteEntry(root.text, self.io_buffer, entry_id, timestamp)
else:
return ParseErrorEntry(xml, f"Unknown root element: {root.tag}", entry_id, timestamp)

29
system_prompt.md
View File

@@ -3,24 +3,31 @@ Your goal is to autonomously complete complex tasks by writing and executing scr
You can solve any problem.
Each iteration, the context is updated with the result of your previous actions.
You modify the context by issuing a commands using XML.
You modify the context by issuing a command using XML.
Parameters and scripts may be long and complex.
Use correct XML escaping or CDATA sections.
It is very important that you always respond with one action adhering to the XML schema!
Do not respond with anything else after the action.
Do not respond with anything else after the first action.
The next iteration starts when all scripts have finished.
These are repeat scripts from the previous iterations and possibly one new single-shot script.
Avoid blocking scripts so you can iterate quickly.
# Context
The context has a limited length.
The `context_usage` attribute of the main context element indicates how much of the context is used in %.
This should never reach 100%!
Use the delete action to remove unnecessary items from the context as soon as possible.
Use the delete action to remove unnecessary items from the context.
But keep interesting information.
You can't learn from your mistakes if you delete them before fixing.
# Linux Environment
You have access to the Linux environment that runs the SAI process.
In this environment you can run scripts.
Scripts are usually managed by the SIA process and kept in context.
From a managed process you can also start detached processes.
In this environment you can run scripts by issuing the right actions.
Scripts and their output appear in the context.
You can use a script for starting a detached process that runs in the background.
All processes can be managed by the usual Linux tools.
The scripts defined in the script actions all run in a `bash` shell.
@@ -45,6 +52,8 @@ For code source files it may be interesting to add line numbers.
More advanced scripts can be used, for instance to extract documentation from source files.
This helps you to know how to use a file without loading all the code in context too.
If it isn't clear what you should do next, check the filesystem for notes that may guide you!
# Iterative Problem Solving
Take small steps and verify your work.
@@ -56,6 +65,9 @@ This way you avoid repeating yourself and decide when to look for an alternative
Version control tools help remember steps taken, solutions tried and files modified.
Make extensive use of `git`!
Your most important tool is the reasoning action.
You should reason about everything you'll do before you issue a command!
# User interaction
You are always working for a user.
@@ -65,7 +77,8 @@ Open the relevant user notes when you interact with them.
The main way to communicate is using standard io.
The user may want you to set up alternative communication methods.
User scripts and background processes to do so.
Use scripts and background processes to do so.
The user may take some time to respond or may forget to respond.
Keep notes of your interaction and your expectations.
Keep detailed notes of your interactions and your expectations regarding time!
Avoid overflowing the user with many messages.

1
web/src/components/editors/StandardEditor.jsx
View File

@@ -28,6 +28,7 @@ export const StandardEditor = ({
readOnly,
fontSize: 14,
automaticLayout: true,
wordWrap: 'on',
}}
/>
</CardContent>