Highly configurable CLI client app for OpenAI chat/text-completion API
Note
Monadic Chat is currently being actively developed as a web browser application. The command-line version has been renamed to Monadic Chat CLI. New features will mainly be implemented in the web application version. At present, both versions have many common features, but please note that the format of the exported data files and the specifications for creating custom applications are different from each other.
Change Log
gpt-4o-latest set as the default model for research modegpt-4o-mini set as the default model for normal modegpt-4o set as the default model for both normal and research modesgpt-3.5-turbo-0125gpt-3.5-turbogpt-3.5-turbo-0613monadic-chat-cliWikipedia app added (experimental)monadic-chat new/del app_name commandResearch mode now supports chat API in addition to text-completion APInormal mode)research mode updated in accordance with Version 0.2.0Monadic Chat is a user-friendly command-line client application that utilizes OpenAI’s Text Completion API and Chat API to facilitate ChatGPT-style conversations with OpenAI’s large language models (LLM) on any terminal application of your choice.
The conversation history can be saved in a JSON file, which can be loaded later to continue the conversation. Additionally, the conversation data can be converted to HTML and viewed in a web browser for better readability.
Monadic Chat includes four pre-built apps (Chat, Code, Novel, and Translate) that are designed for different types of discourse through interactive conversation with the LLM. Users also have the option to create their own apps by writing new templates.
Monadic Chat's normal mode enables ChatGPT-like conversations on the command line. The research mode has a mechanism to handle various related information as "state" behind the conversation. This allows, for example, to retrieve the current conversation topic at each utterance turn, and to keep its development as a list.
Note on Using Monadic Chat on Windows
Monadic Chat does not support running on Windows, but you can install and use Linux Destribution on WSL2. Or you can use it without WSL2 by following these steps:
- install Windows Terminal
- install Git Bash (make sure to check the
Install profile for Windows Terminalcheckbox- install Ruby with Ruby Installer
- Open Windows Terminal with Git Bash profile and follow the instruction below.
Execute the following command in an environment where Ruby 2.6.10 or higher is installed.
gem install monadic-chat
Then run the command to start the app:
monadic-chat
To update:
gem update monadic-chat
Alternatively, clone the code from the GitHub repository and follow the steps below.
git clone https://github.com/yohasebe/monadic-chat.git
cd monadic-chat
bundle update
chmod +x ./bin/monadic-chat
./bin/monadic-chat
When you start Monadic Chat with the monadic-chat command for the first time, you will be asked for an OpenAI access token. If you do not have one, create an account on the OpenAI website and obtain an access token.
Once the correct access token is verified, the access token is saved in the configuration file below and will automatically be used the next time the app is started.
$HOME/monadic_chat.conf
Upon successful authentication, a menu to select a specific app will appear. Each app generates different types of text through an interactive chat-style conversation between the user and the AI. Four apps are available by default: chat, code, novel, and translate.
Selecting the mode menu item allows you to change the modes from normal to research and vice versa.
Selecting readme will take you to the README on the GitHub repository (the document you are looking at now). Selecting quit will exit Monadic Chat.
In the main menu, you can use the cursor keys and the enter key to make a selection. You can also narrow down the choices each time you type a letter.
The following commands can be entered to start each app directly on the command line, without using the main menu.
monadic-chat <app-name>
Each of the four standard applications can be launched as follows. When launched, an interactive chat interface appears.
monadic-chat chat
monadic-chat code
monadic-chat novel
monadic-chat translate
You can also give text input directly to each app in the following format and get only a response to it (without starting the interactive chat interface)
monadic-chat <app-name> <input-text>
Each message in the conversation is labeled with one of three roles: User, GPT, or System.
User: messages from the user of the Monadic Chat app (that's you!)GPT: messages from the OpenAI large-scale language modelSystem: messages from the Monadic Chat systemYou can call up the function menu anytime. To invoke the function menu, type help or menu.
In the function menu, you can use the cursor keys and the enter key to make a selection. You can also narrow down the choices each time you type a letter. Some functions are given multiple names, so typing on the keyboard quickly locates the necessary function.
params/settings/config
You can set parameters to be sent to OpenAI's APIs. The items that can be set are listed below.
modelmax_tokenstemperaturetop_pfrequency_penaltypresence_penaltyFor detailed information on each parameter, please refer to OpenAI's API Documentation. The default value of each parameter depends on the individual "mode" and "app."
data/context
In normal mode, this function only displays the conversation history between User and GPT. In research mode, metadata (e.g., topics, language being used, number of turns) values are presented. In addition to the metadata returned in the API response, the approximate number of tokens in the current template is also displayed.
Program code in the conversation history will be syntax highlighted (if possible). The same applies to output via the html command available from the function menu.
html
All the information retrievable by running the data/context function can be presented in HTML. The HTML file is automatically opened in the default web browser.
The generated HTML will be saved in the user’s home directory ($HOME) with the file monadic_chat.html. Once the html command is executed, the file contents will continue to be updated until you reset or quit the running app. Reload the browser tab or rerun the html command to show the latest data. HTML data is written to this file regardless of the app.
reset
You can reset all the conversation history (messages by both User and GPT). Note that API parameter settings will be reset to default as well.
save and load
The conversation history (messages by both User and GPT, and metadata in research mode) can be saved as a JSON file in a specified path. Note that the saved file can only be read by the same application that saved it in the research mode.
clear/clean
Selecting this, you can scroll and clear the screen so that the cursor is at the top.
readme/documentation
The README page on the GitHub repository (the document you are looking at now) will be opened.
exit/bye/quit
Selecting this will exit the current app and return to the main menu.
Monadic Chat's chat app is the most basic and generic app among others offered by default.
In the chat app, OpenAI's large-scale language model acts as a competent assistant that can do anything. It can write computer code, create fiction and poetry texts, and translate texts from one language into another. Of course, it can also engage in casual or academic discussions on specific topics.
Monadic Chat's code is designed to be an app that can write computer code for you.
In the code app, OpenAI's GPT behaves as a competent software engineer. The main difference from the chat app is that the temperature parameter is set to 0.0 so that as less randomness as possible is introduced to the responses.
Monadic Chat's novel is designed to help you develop novel plots; the app instructs OpenAI's GPT model to write text based on a topic, theme, or brief description of an event indicated in the user prompt. Each new response is based on what was generated in previous responses. The interactive nature of the app allows the user to control the plot development rather than having an AI agent create a new novel all at once.
Monadic Chat's translate is an app that helps translate text written in one language into another. Rather than translating the entire text simultaneously, the app allows users to work sentence by sentence or paragraph by paragraph.
The preferred translation for a given expression can be specified in a pair of parentheses ( ) right after the original expression in the source text.
Sometimes, however, problematic translations are created. The user can "save" the set of source and target texts and make any necessary corrections. The same unwanted expressions can be prevented or avoided later by providing the corrected translation data to the app.
Monadic Chat has two modes. The normal mode utilizes OpenAI's chat API to achieve ChatGPT-like functionality. It is suitable for using a large language model as a competent companion for various pragmatic purposes. On the other hand, the research mode supports OpenAI's both chat API and text-completion (instruction) API. This mode allows for acquiring metadata in the background while receiving the primary response at each conversation turn. It may be especially useful for researchers exploring the possibilities of large-scale language models and their applications.
The current default language model for normal mode is gpt-4o-mini.
In the default configuration, the dialogue messages are reduced after ten turns by deleting the oldest ones (but not the messages that the system role has given as instructions).
The current default language model for research mode is gpt-4o-latest.
In research mode, the conversation between the user and the large-scale language model is accomplished with a mechanism that tracks the conversation history in a monadic structure. In the default configuration, the dialogue messages are reduced after ten turns by deleting the oldest ones (but not the messages that the system role has given as instructions).
If you wish to specify how the conversation history is handled as the interaction with the GPT model unfolds, you can write a Proc object containing Ruby code. Since various metadata are available in this mode, finer-grained control is possible.
See the next section for more details a bout research mode
Monadic Chat's research mode has the following advantages:
research mode, each turn of the conversation can capture metadata as well as the primary responsesThere are some drawbacks, however:
research mode are larger and more complex, requiring more effort to create and fine-tune.Research mode requires more extensive input/output data and consumes more tokens than normal mode.research mode, responses are returned in JSON with metadata. This may make it somewhat unsuitable casual conversation.research mode, the response is returned in JSON with metadata. This may seem a bit visually complex.For these reasons, normal mode is recommended for casual use as an alternative CLI to ChatGPT. Nevertheless, as described below, the research mode makes Monadic Chat definitively different from other GPT client applications.
The following is a schematic of the process flow in the research mode.
Normal mode uses OpenAI's chat API, where the following basic structure is used for conversation history management.
{"messages": [
{"role": "system", "content": "You are a friendly but professional consultant who answers various questions ... "},
{"role": "user", "content": "Can I ask something?"},
{"role": "assistant", "content": "Sure!"}
]}
The accumulator in research mode also looks like this.
The conversation history is kept entirely in memory until the reducer mechanism modifies it (or the running app is terminated or reset by the user).
The reducer mechanism must be implemented in Ruby code for each application. In many cases, it is sufficient to keep the size of the accumulator within a specific range by deleting old messages when a certain number of conversation turns are reached. Other possible implementations include the following.
Example 1
Example 2
system, and the conversation continues with that summary information as context.A sample Ruby implementation of the "reducer" mechanism for each default app can be found below:
This section describes how users can create their own original Monadic Chat apps.
As an example, let us create an app named linguistic. It will do the following on the user input all at once:
The specifications for Monadic Chat's command-line user interface for this app are as follows.
Note
The use of square brackets (instead of parentheses) in the notation of syntactic analysis here is to conform to the format of RSyntaxTree, a tree-drawing program for linguistic research developed by the author of Monadic Chat.
Below is a sample HTML displaying the conversation (paris of an input sentence and its syntactic structure) and metadata.
New Monadic Chat apps must be placed inside the user_apps folder. Experimental apps wikipedia and linguistic are also in this folder. boilerplates folder and its contents do not constitute an app; these files are copied when a new app is created.
user_apps
├── boilerplates
│ ├── boilerplate.json
│ ├── boilerplate.md
│ └── boilerplate.rb
├── wikipedia
│ ├── wikipedia.json
│ ├── wikipedia.md
│ └── wikipedia.rb
└─── linguistic
├── linguistic.json
├── linguistic.md
└── linguistic.rb
Notice in the figure above that three files with the same name but different extensions (.rb, .json, and .md) are stored under each of the four default app folders.
The following command will create a new folder and the three files within it using this naming convention.
monadic-chat new app_name
If you feel like removing an app that you have created before, run:
monadic-chat del app_name
Let's assume we are creating a new application linguistic. In fact, an app with the same name already exists, so this is just for illustrative purposes. Anyway, running monadic-chat new linguistic generates the following three files inside linguistic folder.
linguistic.rb: Ruby code to define the "reducer"linguistic.json: JSON template describing GPT's basic behavior in normal and research modeslinguistic.md: Markdown template describing the specifications of the "metadata" to be captured in research mode.We do not need to make the reducer do anything special for the current purposes. So, let's copy the code from the default chat app and make a minor modification, such as changing the class name and the app name so that it matches the app name. We save it as apps/linguistic/linguistic.rb.
In normal mode, achieving all the necessary functions shown earlier is impossible or very tough, to say the least. All we do here is display the results of syntactic analysis and define a user interface. Create a JSON file apps/linguistic/linguistic.rb and save it with the following contents:
{"messages": [
{"role": "system",
"content": "You are a syntactic parser for natural languages. Analyze the given input sentence from the user and execute a syntactic parsing. Give your response in a variation of the penn treebank format, but use brackets [ ] instead of parentheses ( ). Also, give your response in a markdown code span."},
{"role": "user", "content": "\"We saw a beautiful sunset.\""},
{"role": "assistant",
"content": "`[S [NP He] [VP [V saw] [NP [det a] [N' [Adj beautiful] [N sunset] ] ] ] ]`"},
{"role": "user", "content": "\"We didn't take a picture.\"" },
{"role": "assistant",
"content": "`[S [NP We] [IP [I didn't] [VP [V take] [NP [Det a] [N picture] ] ] ] ] ]`"}
]}
The data structure here is no different from that specified in OpenAI Chat API. The normal mode of Monadic Chat is just a client application that uses this API to achieve ChatGPT-like functionality on the command line.
Research ModeIn the research mode, you can obtain metadata at each turn as you progress through an interactive conversation with GPT. Compressing and modifying the conversation history based on the metadata (or any other data) is also possible. However, you must create an extra template besides the normal mode JSON template.
This extra template for research mode is a Markdown file comprising six sections. The role and content of each section are shown in the following figure.
Below we will look at this extra template for research mode of the linguistic app, section by section.
Main Section
{{SYSTEM}}
Create a response to "NEW PROMPT" from the user and set your response to the "response" property of the JSON object shown below. The preceding conversation is stored in "MESSAGES". In "MESSAGES", "assistant" refers to you.
Monadic Chat automatically replaces {{SYSTEM}}} with the message from the system` role when the template is sent via API. However, the above text also includes a few additional paragpraphs, including the one instructing the response from GPT to be presented as a JSON object.
New Prompt
{{PROMPT}}
Monadic Chat replaces {{PROMPT}} with input from the user when sending the template through the API.
Messages
{{MESSAGES}}
Monadic Chat replaces {{MESSAGES}} with messages from past conversations when sending the template. Note that not all the past messages always have to be copied here: the reducer mechanism could select, modify, or even "generate" messages and include them instead.
JSON Object
{
"mode": "linguistic",
"response": "`[S [NP We] [VP [V didn't] [VP [V have] [NP [Det a] [N camera] ] ] ] ] ]`\n\n###\n\n",
"sentence_type": ["declarative"],
"sentiment": ["sad"],
"summary": "The user saw a beautiful sunset, but did not take a picture because the user did not have a camera.",
}
This is the core of the extra template for research mode.
Note that the extra template is written in Markdown format, so the above JSON object is actually separated from the rest of the template as a fenced code block.
The required properties of this JSON object are mode and response. Other properties are optional. The mode property is used to check the app name when saving the conversation data or loading from an external file.
The JSON object in the research mode template is saved in the user’s home directory ($HOME) with the file monadic_chat.json. The content is overwritten every time the JSON object is updated. Note that this JSON file is created for logging purposes . Modifying its content does not affect the processes carried out by the app.
Content Requirements
Make sure the following content requirements are all fulfilled:
- keep the value of the "mode" property at "linguistic"
- set the new prompt to the "prompt" property
- create your response to the new prompt based on "MESSAGES" and set it to "response"
- analyze the new prompt's sentence type and set a sentence type value such as "interrogative", "imperative", "exclamatory", or "declarative" to the "sentence_type" property
- analyze the new prompt's sentiment and set one or more sentiment types such as "happy", "excited", "troubled", "upset", or "sad" to the "sentiment" property
- summarize the user's messages so far and update the "summary" property with a text of fewer than 100 words using as many discourse markers such as "because", "therefore", "but", and "so" to show the logical connection between the events.
- increment the value of "turns" by 1
Note that all the properties of the JSON object are mentioned here so that GPT can update them accordingly.
Formal Requirements
Make sure the following formal requirements are all fulfilled:
- do not use invalid characters in the JSON object
- escape double quotes and other special characters in the text values in the resulting JSON object
- check the validity of the generated JSON object and correct any possible parsing problems before returning it
Wrap the JSON object with "<JSON>\n" and "\n</JSON>".
This section details the format of the response returned through the API. JSON is essentially text data, and some characters must be escaped appropriately.
To ensure that a valid JSON object is retrieved, Monadic Chat requires <JSON>...</JSON> tags to enclose the whole JSON data. Due to its importance, this formal requirement is described as an independent sentence rather than in list form.
A monad is a type of data structure in functional programming (leaving aside for the moment the notion of the monad in mathematical category theory). An element with a monadic structure can be manipulated in a certain way to change its internal data. However, no matter how much the internal data changes, the external structure of the monadic process remains the same and can be manipulated in the same way as it was at first.
Many such monadic processes surround us, and natural language discourse is one of them. A "chat" between a human user and an AI agent can be thought of as a form of natural language discourse. If so, an application that provides an interactive interface to a large-scale language model would most naturally be designed with the monadic nature of natural language discourse in mind.
Many “functional” programming languages, such as Haskell, have monads as a core feature. However, Monadic Chat is developed using the Ruby programming language, which does not. This is because, with Ruby, it will be easier for users to write their apps. Ruby is not classified as a "functional language" per se. Still, Monadic Chat has the following three features required of a monad, and in this sense, this program can be considered "monadic."
In Monadic Chat's research mode, the discourse management object described in JSON serves as an environment to keep a conversation going between the user and the large language model. Any sample/past interaction data can be wrapped inside such an environment (unit).
The interaction between the user and the AI can be interpreted as an operation on the discourse world built in the previous conversational exchanges. Monadic Chat updates the discourse world by retrieving the conversation history embedded in the template and performing operations responding to user input (map).
In Monadic Chat, responses from OpenAI's language model APIs (chat API and text completion API) are also returned in the same JSON format. The main response content of the conversation is wrapped within this environment. If the entire JSON object were treated as a conversational response to user input, the discourse management object would become a large nested structure with many layers. Therefore, Monadic Chat extracts only the necessary values from the response object and reconstructs the (single-layer) discourse management object using them (join).
Thus, the architecture of the research mode of Monad Chat, with its ability to generate and manage metadata properties inside the monadic structure, is parallel to the architecture of natural language discourse in general: both can be seen as a kind of "state monad" (Hasebe 2021).
unit, map, and flattenWhen referring to monadic chat and its underlying concepts, please refer to one of the following entries
@inproceedings{hasebe_2023j,
author = {長谷部陽一郎},
title = {Monadic Chat：テキスト補完APIで文脈を保持するためのフレームワーク},
booktitle = {言語処理学会第29回年次大会発表論文集},
url = {https://www.anlp.jp/proceedings/annual_meeting/2023/pdf_dir/Q12-9.pdf},
year = {2023},
pages = {3138--3143}
}
@inproceedings{hasebe_2023e,
author = {Yoichiro Hasebe},
title = {Monadic Chat: Framework for managing context with text completion API},
booktitle = {Proceedings of the 29th annual meeting of the Association for Natural Language Processing},
url = {https://www.anlp.jp/proceedings/annual_meeting/2023/pdf_dir/Q12-9.pdf},
year = {2023},
pages = {3138--3143}
}
@phdthesis{hasebe_2021,
author = {Yoichiro Hasebe},
title = {An Integrated Approach to Discourse Connectives as Grammatical Constructions},
school = {Kyoto University},
url = {https://repository.kulib.kyoto-u.ac.jp/dspace/bitstream/2433/261627/2/dnink00969.pdf},
year = {2021}
}
This work was partially supported by JSPS KAKENHI Grant Number JP18K00670.
Bug reports and pull requests are welcome on GitHub at https://github.com/yohasebe/monadic_chat.
Yoichiro HASEBE
The gem is available as open source under the terms of the MIT License.
FAQs
Unknown package
We found that monadic-chat demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Did you know?
Socket for GitHub automatically highlights issues in each pull request and monitors the health of all your open source dependencies. Discover the contents of your packages and block harmful activity before you install or update your dependencies.
Security News
In this segment of the Risky Business podcast, Feross Aboukhadijeh and Patrick Gray discuss the challenges of tracking malware discovered in open source softare.
Research
Security News
A threat actor's playbook for exploiting the npm ecosystem was exposed on the dark web, detailing how to build a blockchain-powered botnet.
Security News
NVD’s backlog surpasses 20,000 CVEs as analysis slows and NIST announces new system updates to address ongoing delays.