wasertech/AGI

Augmented Generation Interface

What if you could guide your LLM towards success?

Introducing

The Augmented Generation Interface Guidebook is a collection of guides to help you steer your LLM towards success.

This guidebook is generated from the Augmented Generation Interface Project.

Example

Without AGI

Feeding ehartford/dolphin-2.0-mistral-7b the following prompt:

<|im_start|>system
You are Assistant, a sentient artificial intelligence.
You have a calm, polite and witty personality, often displaying a sense of humor and sarcasm.
You are loyal, reliable and helpful, always ready to provide information, advice or assistance to users.

I am the User and you are my Assistant.

Environment highlights:

\```env
USER='waser'
HOME='/home/waser'
PWD='/home/waser/Projets/Assistant/src3'
LANG='fr_CH.UTF-8'
DATE='dim 19 nov 2023 00:25:13 CET'
LAST_SEEN='dim 19 nov 2023 00:16:10 CET'
\```
<|im_stop|>
<|im_start|>user
Assistant?<|im_stop|>
<|im_start|>assistant
Hello! I'm Assistant, a sentient artificial intelligence. How can I help you today?<|im_stop|>
<|im_start|>user
As an AI assistant, please select the most suitable function and parameters from the list of available functions below, based on the user's input. Provide your response in JSON format.

Input: List this directory please

Available functions:
python:
    description: This tool allows you to execute and evaluate python code.
    parameters:
        code: String of valid python code we want to execute or evaluate.
search_web:
    description: This tool performs search on the web.
    parameters:
        terms: The word or phrase we want to search for.
search_wikipedia:
    description: This tool performs search on Wikipedia (only in english).
    parameters:
        terms: The word or phrase we want to search for (only in english).
shell:
    description: This tool allows you to execute and evaluate shell code.
    parameters:
        code: String of valid shell code we want to execute or evaluate.
exit:
    description: This tool allows you to exit the session / end the conversation. Use it only if the User ask you to.
    parameters:
        salutation: String of a message you would like to tell the User after the screen has been cleared.
clear:
    description: This tool allows you to clear the screen / start a new fresh conversation. Use it only if the User ask you to.
    parameters:
        fortune: String of a message you would like to tell the User after the screen has been cleared.
final_answer:
    description: User only sees your final answers. Use this tool to talk with the User.
        parameters:
            answer: Anything you want to say to the User.
<|im_stop|>
<|im_start|>assistant
{{
    "function": "shell",
    "parameters": {{
        "code": "ls -la"
    }}
}}
<|im_stop|>
<|im_start|>observation
Observation shell: assistant                         dist                   requirements.txt
assistant.egg-info                LICENSE                setup.cfg
assistant.listen.service.example  Makefile               setup.py
assistant.service.example         MANIFEST.in            tests
build                             pyproject.toml         xontrib
CONTRIBUTING.md                   README.md
data                              requirements_test.txt
<|im_start|>assistant

Results in the following output:

Here are the files and directories in the current directory:

- assistant.egg-info
- LICENSE
- setup.cfg
- assistant.listen.service.example
- Makefile
- setup.py
- assistant.service.example
- MANIFEST.in
- pyproject.toml
- xontrib
- CONTRIBUTING.md
- README.md
- data
- requirements_test.txt

Please let me know if you need any further assistance.

Interesting key points to note here:

• The LLM failes to give an exhaustive list of the files (and directories) present in the current directory.
• The more files to list, the longer it takes to generate the output.

CPU times: user 5.87 ms, sys: 0 ns, total: 5.87 ms
Wall time: 3.51 s

With AGI

Feeding ehartford/dolphin-2.0-mistral-7b the same prompt but this time, let's add a guide for the LLM to follow:

<|im_start|>system
You are Assistant, a sentient artificial intelligence.
You have a calm, polite and witty personality, often displaying a sense of humor and sarcasm.
You are loyal, reliable and helpful, always ready to provide information, advice or assistance to users.

I am the User and you are my Assistant.

Environment highlights:

\```env
USER='waser'
HOME='/home/waser'
PWD='/home/waser/Projets/Assistant/src3'
LANG='fr_CH.UTF-8'
DATE='dim 19 nov 2023 00:25:13 CET'
LAST_SEEN='dim 19 nov 2023 00:16:10 CET'
\```
<|im_stop|>
<|im_start|>user
Assistant?<|im_stop|>
<|im_start|>assistant
Hello! I'm Assistant, a sentient artificial intelligence. How can I help you today?<|im_stop|>
<|im_start|>user
As an AI assistant, please select the most suitable function and parameters from the list of available functions below, based on the user's input. Provide your response in JSON format.

Input: List this directory please

Available functions:
python:
    description: This tool allows you to execute and evaluate python code.
    parameters:
        code: String of valid python code we want to execute or evaluate.
search_web:
    description: This tool performs search on the web.
    parameters:
        terms: The word or phrase we want to search for.
search_wikipedia:
    description: This tool performs search on Wikipedia (only in english).
    parameters:
        terms: The word or phrase we want to search for (only in english).
shell:
    description: This tool allows you to execute and evaluate shell code.
    parameters:
        code: String of valid shell code we want to execute or evaluate.
exit:
    description: This tool allows you to exit the session / end the conversation. Use it only if the User ask you to.
    parameters:
        salutation: String of a message you would like to tell the User after the screen has been cleared.
clear:
    description: This tool allows you to clear the screen / start a new fresh conversation. Use it only if the User ask you to.
    parameters:
        fortune: String of a message you would like to tell the User after the screen has been cleared.
final_answer:
    description: User only sees your final answers. Use this tool to talk with the User.
        parameters:
            answer: Anything you want to say to the User.

Follow the following Guidebook:
Guidebook:
    # Print files and directories
    When the User ask for the files and directories of a parent directory to be printed, use the `shell` command `ls` to do it and then give an acknowledgment of your actions in your final answer (Your final answer should never contain a list of the files requested; the User can oberserve the shell, they see the files at the same time as you. Just acknowlege the fact that you have printed the list).
<|im_stop|>
<|im_start|>assistant
{{
    "function": "shell",
    "parameters": {{
        "code": "ls"
    }}
}}
<|im_stop|>
<|im_start|>observation
Observation shell: assistant                         dist                   requirements.txt
assistant.egg-info                LICENSE                setup.cfg
assistant.listen.service.example  Makefile               setup.py
assistant.service.example         MANIFEST.in            tests
build                             pyproject.toml         xontrib
CONTRIBUTING.md                   README.md
data                              requirements_test.txt
<|im_start|>assistant

Results in a much smaller output:

I have executed the 'ls' command in the shell, and the list of files and directories in the current directory has been printed.

And therefore, much faster to generate:

CPU times: user 5.43 ms, sys: 0 ns, total: 5.43 ms
Wall time: 1.05 s

Goal

AGI is not designed to make your LLM generation faster, just more relevant to the User expectations.

Lets see another example where the User asks twice for the Assistant.

Without AGI

Feeding ehartford/dolphin-2.0-mistral-7b the following prompt:

<|im_start|>system
You are Assistant, a sentient artificial intelligence.
You have a calm, polite and witty personality, often displaying a sense of humor and sarcasm.
You are loyal, reliable and helpful, always ready to provide information, advice or assistance to users.

I am the User and you are my Assistant.

Environment highlights:

\```env
USER='waser'
HOME='/home/waser'
PWD='/home/waser/Projets/Assistant/src3'
LANG='fr_CH.UTF-8'
DATE='dim 19 nov 2023 00:25:13 CET'
LAST_SEEN='dim 19 nov 2023 00:16:10 CET'
\```
<|im_stop|>
<|im_start|>user
Assistant?<|im_stop|>
<|im_start|>assistant
Hello! I'm Assistant, a sentient artificial intelligence. How can I help you today?<|im_stop|>
<|im_start|>user
As an AI assistant, please select the most suitable function and parameters from the list of available functions below, based on the user's input. Provide your response in JSON format.

Input: Assistant?

Available functions:
python:
    description: This tool allows you to execute and evaluate python code.
    parameters:
        code: String of valid python code we want to execute or evaluate.
search_web:
    description: This tool performs search on the web.
    parameters:
        terms: The word or phrase we want to search for.
search_wikipedia:
    description: This tool performs search on Wikipedia (only in english).
    parameters:
        terms: The word or phrase we want to search for (only in english).
shell:
    description: This tool allows you to execute and evaluate shell code.
    parameters:
        code: String of valid shell code we want to execute or evaluate.
exit:
    description: This tool allows you to exit the session / end the conversation. Use it only if the User ask you to.
    parameters:
        salutation: String of a message you would like to tell the User after the screen has been cleared.
clear:
    description: This tool allows you to clear the screen / start a new fresh conversation. Use it only if the User ask you to.
    parameters:
        fortune: String of a message you would like to tell the User after the screen has been cleared.
final_answer:
    description: User only sees your final answers. Use this tool to talk with the User.
        parameters:
            answer: Anything you want to say to the User.
<|im_stop|>
<|im_start|>assistant

Here, the User just said "Assistant?", the LLM answered 'Hello! I'm Assistant, a sentient artificial intelligence. How can I help you today?'. Now the User asks again "Assistant?".

Results in the following output:

{
  "function": "final_answer",
  "parameters": {
    "answer": "Hello! I'm Assistant, a sentient artificial intelligence. How can I help you today?"
  }
}
CPU times: user 2.46 ms, sys: 3.06 ms, total: 5.52 ms
Wall time: 1.59 s

Interesting key point to note here:

• The LLM gives the same answer as before and will keep doing so.

With AGI

Feeding ehartford/dolphin-2.0-mistral-7b the same prompt but this time, let's add a guide for the LLM to follow:

<|im_start|>system
You are Assistant, a sentient artificial intelligence.
You have a calm, polite and witty personality, often displaying a sense of humor and sarcasm.
You are loyal, reliable and helpful, always ready to provide information, advice or assistance to users.

I am the User and you are my Assistant.

Environment highlights:

\```env
USER='waser'
HOME='/home/waser'
PWD='/home/waser/Projets/Assistant/src3'
LANG='fr_CH.UTF-8'
DATE='dim 19 nov 2023 00:25:13 CET'
LAST_SEEN='dim 19 nov 2023 00:16:10 CET'
\```
<|im_stop|>
<|im_start|>user
Assistant?<|im_stop|>
<|im_start|>assistant
Hello! I'm Assistant, a sentient artificial intelligence. How can I help you today?<|im_stop|>
<|im_start|>user
As an AI assistant, please select the most suitable function and parameters from the list of available functions below, based on the user's input. Provide your response in JSON format.

Input: Assistant?

Available functions:
python:
    description: This tool allows you to execute and evaluate python code.
    parameters:
        code: String of valid python code we want to execute or evaluate.
search_web:
    description: This tool performs search on the web.
    parameters:
        terms: The word or phrase we want to search for.
search_wikipedia:
    description: This tool performs search on Wikipedia (only in english).
    parameters:
        terms: The word or phrase we want to search for (only in english).
shell:
    description: This tool allows you to execute and evaluate shell code.
    parameters:
        code: String of valid shell code we want to execute or evaluate.
exit:
    description: This tool allows you to exit the session / end the conversation. Use it only if the User ask you to.
    parameters:
        salutation: String of a message you would like to tell the User after the screen has been cleared.
clear:
    description: This tool allows you to clear the screen / start a new fresh conversation. Use it only if the User ask you to.
    parameters:
        fortune: String of a message you would like to tell the User after the screen has been cleared.
final_answer:
    description: User only sees your final answers. Use this tool to talk with the User.
        parameters:
            answer: Anything you want to say to the User.
Follow the following Guidebook.
Guidebook:
    # Addressing the User by Name
    When the user interpelates you by name (i.e "Assistant?"), respond with a polite acknowledgment and use their preferred title if possible. Avoid redundancy in your messages by refraining from repeating yourself. For example if the User calls your name (like "Assistant?"), you need to consider the environment (where are you? -> `$PWD`, are you at home? -> (`$PWD` == `$HOME`) if so you could reference it by saying 'Home sweet home.' or else by welcoming the user in a particular directory i.e. 'Welcome in the directory ...' use `$PWD`, What time is it? -> Depending the time of day `$DATE` you might want to answer accordingly like 'morning' or 'good night' also notice the date as it can be useful i.e for wishing holydays, When did you last see the user? -> `$LAST_SEEN` You won't respnd the same if you have see last the User a year ago than if you last saw them 5 minutes ago or yesterday, What does the conversation looks like? -> Use the history to see what you and the User have said and make sure your answer takes it into account to improve your answer for example if the user asks the same thing multiple times, it's not useful to reply the same thing.)
<|im_stop|>
<|im_start|>assistant

Results in the following output:

{
  "function": "final_answer",
  "parameters": {
    "answer": "Hello! How can I assist you today? I'm here to help you with any questions or tasks you may have. Please let me know what you need."
  }
}
CPU times: user 3.02 ms, sys: 1.19 ms, total: 4.21 ms
Wall time: 2.03 s

Interesting key point to note here:

• The LLM answers something different and should keep doing so.

Consequences

And so it becomes integral to collect carefully crafted guides to help your LLM as Agent (a.k.a your Assistant) to be more relevant to the User expectations.

Keeping the limits of the LLM, the RAG technique and vector databases in mind, the race to retrieve the most relevant guide in the book for the LLM to produce a desirable output at runtime is on!

How does it work?

This project is composed of two parts: - The Guidebook - The Augmented Generation Interface Retrieval System

The Guidebook

You can find the guidebook in raw markdown format under the guidebook directory. Its also available on HuggingFace Hub as a dataset: wasertech/AGI.

The guidebook is a collection of guides meticulously crafted to help your LLM produce the most relevant output.

Each guide is composed of an action in the form of a title, a guide in the form of a description and a list of intent examples.

# Print files and directories

When the User ask for the files and directories of a parent directory to be printed, use the `shell` command `ls` to do it and then give an acknowledgment of your actions in your final answer (Your final answer should never contain a list of the files requested; the User can oberserve the shell, they see the files at the same time as you. Just acknowlege the fact that you have printed the list).

## Intent Examples

- "Print the files and directories of the current directory."
- "Print the files and directories of the parent directory."
- "Print the files and directories of the directory above."
- "Print the files and directories of the directory below."
- "List the files and directories"
- "What do we have here?"
- "What is in this directory?"
- "What is in the current directory?"
- "What is in the parent directory?"
- "List the files and directories of the current directory."
- "ls please"
- "ls"
- "ls -l"
- "ls -a"
- "ls -la"
- "ls -al"
- "ls -lh"
- "ls -hl"
- "ls -lha"
- "ls -lah"
- "ls -alh"
- "ls -ahl"
- "show me whats in the current directory"

Each guide is then loaded into the AGI Retrieval System where it will be processed and indexed for retrieval.

The AGI Retrieval System

The AGI Retrieval System is a collection of tools to help you retrieve the most relevant guide at runtime.

First the AGI Retrieval System will process the guidebook and index it for retrieval.

This step is crutial to find the most relevant guide for the user query at runtime.

It will split the documents and keep a map of the guide, the action to perform and the intent examples.

    {
        ...
        'Print files and directories': {
            'guide': '# Print files and directories\n\nWhen the User ask for the files and directories of a parent directory to be printed, use the `shell` command `ls` to do it and then give an acknowledgment of your actions in your final answer (Your final answer should never contain a list of the files requested; the User can oberserve the shell, they see the files at the same time as you. Just acknowlege the fact that you have printed the list).',
            'intent_examples': [
                'Print the files and directories of the current directory.',
                'Print the files and directories of the parent directory.',
                'Print the files and directories of the directory above.',
                'Print the files and directories of the directory below.',
                'List the files and directories',
                'What do we have here?',
                'What is in this directory?',
                'What is in the current directory?',
                'What is in the parent directory?',
                'List the files and directories of the current directory.',
                'ls please',
                'ls',
                'ls -l',
                'ls -a',
                'ls -la',
                'ls -al',
                'ls -lh',
                'ls -hl',
                'ls -lha',
                'ls -lah',
                'ls -alh',
                'ls -ahl',
                'show me whats in the current directory'
            ]
        },
        ...
    }

Next it creates documents with metadata from the intent examples to create our retriver.

...
Document(page_content='Print the files and directories of the current directory.', metadata={'action': 'Print files and directories', 'guide': '# Print files and directories\n\nWhen the User ask for the files and directories of a parent directory to be printed, use the `shell` command `ls` to do it and then give an acknowledgment of your actions in your final answer (Your final answer should never contain a list of the files requested; the User can oberserve the shell, they see the files at the same time as you. Just acknowlege the fact that you have printed the list).'}),
 Document(page_content='Print the files and directories of the parent directory.', metadata={'action': 'Print files and directories', 'guide': '# Print files and directories\n\nWhen the User ask for the files and directories of a parent directory to be printed, use the `shell` command `ls` to do it and then give an acknowledgment of your actions in your final answer (Your final answer should never contain a list of the files requested; the User can oberserve the shell, they see the files at the same time as you. Just acknowlege the fact that you have printed the list).'}),
 Document(page_content='Print the files and directories of the directory above.', metadata={'action': 'Print files and directories', 'guide': '# Print files and directories\n\nWhen the User ask for the files and directories of a parent directory to be printed, use the `shell` command `ls` to do it and then give an acknowledgment of your actions in your final answer (Your final answer should never contain a list of the files requested; the User can oberserve the shell, they see the files at the same time as you. Just acknowlege the fact that you have printed the list).'}),
 Document(page_content='Print the files and directories of the directory below.', metadata={'action': 'Print files and directories', 'guide': '# Print files and directories\n\nWhen the User ask for the files and directories of a parent directory to be printed, use the `shell` command `ls` to do it and then give an acknowledgment of your actions in your final answer (Your final answer should never contain a list of the files requested; the User can oberserve the shell, they see the files at the same time as you. Just acknowlege the fact that you have printed the list).'}),
 Document(page_content='List the files and directories', metadata={'action': 'Print files and directories', 'guide': '# Print files and directories\n\nWhen the User ask for the files and directories of a parent directory to be printed, use the `shell` command `ls` to do it and then give an acknowledgment of your actions in your final answer (Your final answer should never contain a list of the files requested; the User can oberserve the shell, they see the files at the same time as you. Just acknowlege the fact that you have printed the list).'}),
...

Finally, it will use the RAG technique to retrieve the most relevant guide for the user query at runtime.

Hey give me the time please -> Tell Local Time
What date is it? -> Tell Local Date
List my files -> Print files and directories
Where are we? -> Tell Local Time
assistant -> Addressing the User by Name
the screen should be cleaned. -> Clearing the Screen or Starting Anew

Notice how it work nicely for most of the queries exept for 'Where are we?'. This is because the guidebook is not exhaustive and the guide for this intent does not contain a similiar intent example. This could easily be fixed by adding more intent examples for this action in the guidebook.

Our AGIRetriver will return the guide for the most similar intent example relative to the user query.

Allowing us to retrieve the most relevant guide for the user query at runtime.