prompter.vim

Use vim as a tool for efficiently design and run, debug and save your Large Language Models (LLMs) prompts.

Prompter.vim transforms the Vim editor into an efficient prompt engineering environment, effectively replacing LLM proprietary providers web playgrounds like: Azure OpenAI Service Playground or OpenAI Playground.
From version 0.2, the plugin uses LiteLLM as a LLM provider abstraction layer.
LiteLLM calls all LLM APIs using the OpenAI format: Bedrock, Azure, OpenAI, Cohere, Anthropic, Ollama, Sagemaker, HuggingFace, Replicate (100+ LLMs). So you can use prompter.vim with a vast list of different LLM providers!

How it works

Install and set your variable environment to configure model and settings
Press <F9> (key shortcut for :PrompterSetup)
Edit your prompt
Press <F12> (key shortcut for :PrompterGenerate) to get the LLM completion
Goto step 3 to maybe review your prompt
Save your prompt (and completions) into a snapshot text file (e.g. myexperiment.prompt)


prompter.vim in action: editing/interacting with a prompt that simulate a phone conversation (with a technique below described as "dialogues as part of the text prompt")

⚠ A Tool for Prompt Engineers, Not Coders!

Prompter.vim is not primarily designed as a code completion tool, although you can use it also for that purpose.
Instead, this plugin aims to be a general-purpose replacement for web text completion playgrounds, intended for prompt engineers who want to test and debug natural language prompts.

Features

Use almost all available LLMs
Using LiteLLM as a LLM abstraction layer, you can use a huge list of different LLM providers supported by LiteLLM.
Instant LLM Completion
trigger LLM completions with a simple keystroke.
Run-time statistics
Measure completions in terms of latency, used tokens, throughput, etc.
Exploit all vim editor commodities (other plugins, etc.)
Generate prompts inside the editor and seamlessly save all your work to local files.
Completion Highlight
Support basic completions color highlight

Backstory

The idea emerged in the Spring 2023 while I was writing LLM prompts and experimenting with prompt engineering techniques. I was using a straightforward "text completion" approach, where you input your text prompt corpus and request a completion from a Large Language Model (LLM).

My initial approach was to utilize the web playgrounds offered by LLM providers. However, I encountered numerous issues especially while interacting with Azure OpenAI web playgrounds.

For reasons I do not yet comprehend, the web interaction on the Azure web playground slow down considerably after a certain point. I suspect a bug within the completion boxes. Furthermore, I am not fond of the Azure web interface for the "chat completion" mode. A total mess! Instead, the original OpenAI playground is better implemented, and I did not encounter the aforementioned issues.

Nevertheless, both web playgrounds mentioned, permit only one prompt per browser tab. Therefore, when dealing with multiple active prompts (developing a composite application composed of nested/chained template prompts), you must maintain multiple playgrounds open in distinct tabs. When you achieve certain (intermediate) noteworthy outcomes, you must copy all text boxes and save them in versioned files.

Undertaking all of this with web playgrounds is a cumbersome and error-prone process. The final thought was: what if I could run my completion directly inside my vim editor?

🙄 `text` or `chat` completion?

There are two common "completion modes" foreseen in OpenAI or similar current LLMs:

text completion

Completion mode set as text means that LLM completes, the given context window prompt text with a completion text (text in -> text out). An example of such a model setting is the text-da-vinci-003 OpenAI model. To use a text completion mode, the model must support that mode through a specific API.

                  ┌────────────────────────┐
              ┌─  │                        │ ─┐
   context    │   │ bla bla bla            │  │
   window     │   │ bla bla                │  │
      =       │   │ bla bla bla bla        │  │ prompt
    prompt    │   │ bla                    │  │
      +       │   │ bla bla                │  │
  completion  │   │                        │ ─┘
              │   └────────────────────────┘
              │               |
              │         LLM generation
              │               |
              │               v
              │   ┌────────────────────────┐
              │   │                        │ ─┐
              │   │ bla bla                │  │ 
              │   │ bla bla bla            │  │ text completion
              │   │ bla                    │  │
              └─  │                        │ ─┘
                  └────────────────────────┘

chat completion

Completion mode set as chat means that LLM s fine-tuned for chat "roles" (user say, assistant say, ...). Fore details, please read this. The context window prompt is in fact made by

a "system prompt"
a list of "user" and "assistant" messages. An example of such a model setting is the gpt3.5-turbo OpenAI model. To use a chat completion mode, the model must support that mode, trough specific API.

                  ┌────────────────────────┐
              ┌─  │ bla bla bla bla        │ ─┐
              │   │ bla bla bla            │  │
              │   │ bla bla                │  │ system
              │   │ bla bla bla bla        │  │ prompt
   context    │   │ bla                    │  │
   window     │   │ bla bla                │ ─┘
      =       │   └────────────────────────┘
system prompt │   ┌────────────────────────┐
      +       │   │ user: blablabla        │ ─┐
    chat      │   ├────────────────────────┤  │
      +       │   │ assistant: bla bla bla │  │
  completion  │   ├────────────────────────┤  │ chat
              │   │ user: bla bla bla      │  │ prompt
              │   ├────────────────────────┤  │
              │   │ assistant: blabla bla  │  │
              │   ├────────────────────────┤  │
              │   │ user: blabla bla       │ ─┘
              │   └────────────────────────┘
              │               |
              │         LLM generation
              │               |
              │               v
              │   ┌────────────────────────┐
              │   │                        │ ─┐
              └─  │ assistant: bla bla bla │  │ chat completion
                  │                        │ ─┘
                  └────────────────────────┘

⚠️ Warning
Prompter.vim plugin is conceived to work as text completer fast prototyping playground, avoiding the complications of the chat roles.

So if a model that works only in chat mode (e.g. OpenAI GPT3.5-Turbo) is used, behind the scenes (through a LiteLLM text_completion() method) the editor text content (the prompt) is inserted as "system" role prompt. See also: discussion.

I'm aware that using a chat-based model as a text-based model, as described above, is not the optimal usage, but it's a compromise between the simplicity of having a single text completion playground and the complexity of managing chat roles.

📦 Install

This plugin is made in Python3. Check if your vim installation support Python3:
```
vim --version | grep "+python3" | awk '{print $3}'
```
In my case I got +python3. That's the main prerequisite.
Check also which is the Python version which is compiled vim. Extract the precise Python version with command:
```
  vim --version | grep -o -P '(?<=/python)[0-9]+\.[0-9]+'
```
In my case I got 3.8.

Note that vim can ONLY use the Python version (and related packages) wich is compiled. By example if your system Python current version (python3 --version == Python 3.11.6) differs from the vim python version, say Python 3.8, remember vim will see ONLY Python 3.8 packages. To use Python 3.11.6 packages, you must recompile vim.
Install Python package litellm. You must install litellm using pip of the corresponding Python version, e.g. pip3.8.
```
pip3.8 install -U litellm 
```
Install the plugin using your preferred plugin manager, e.g. using vim-plug plug-in manager, insert in your .vimrc file:
```
Plug 'solyarisoftware/prompter.vim'
```

🔑 Environment Variables Setup

# PROVIDER DEPENDENT SETTINGS USING LiteLLM CONFIGURATION 
# https://docs.litellm.ai/docs/providers
# https://docs.litellm.ai/docs/providers/azure

# LLM PROVIDER MANDATORY SETTINGS
export AZURE_API_VERSION=2023-09-01-preview
export AZURE_API_BASE="https://XXXXXXXXXXXXXXX.openai.azure.com/"
export AZURE_API_KEY="YYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYY"

export MODEL="azure/your_deployment-name"

# LLM PARAMETERS OPTIONAL SETTINGS
# translated OpenAI model parameters 
# https://docs.litellm.ai/docs/completion/input#translated-openai-params
export TEMPERATURE=0.3
export MAX_TOKENS=3000
export OPENAI_STOP=""

💡 A good idea is to edit and keep all variables above in a hidden file, e.g. vi ~/.prompter_azure.vim, and execute it with source ~/.prompter_azure.vim

👊 Commands

You can run commands in vim command mode (:) or the associated key:

Command	Key 🚀	Action
`:PrompterSetup`	`<F9>`	Model setup and initial configurations
`:PrompterGenerate`	`<F12>`	Run the LLM text completion
`:PrompterInfo`	`<F10>`	Report current configuration

`:PrompterSetup`

When you enter vim, to activate the Prompter playground environment, first of all run in command mode:

:PrompterSetup

Following the environment settings, if successful, the command print in the status line the model configurations:

Model: azure/gpt-35-turbo completion mode: chat temperature: 0.7 max_tokens: 100

Explanation of values in the status line report:

                               temperature preset value ───────────────────────────┐
                                                                                   │
                                max_tokens preset value ──────────┐                │
                                                                  │                │
      ┌─────┐ ┌────────────┐                 ┌────┐             ┌─┴─┐            ┌─┴─┐
Model:│azure│/│gpt-35-turbo│ completion mode:│chat│ temperature:│0.7│ max_tokens:│100│
      └──┬──┘ └─────┬──────┘                 └──┬─┘             └───┘            └───┘
         │          │                           │
         │          │                           └─ chat or text, depending on the model
         │          │
         │          └── name of the Azure deployment
         │
         └───────────── name of the LLM provider

`:PrompterGenerate`

Edit your prompt on a vim windows, and to run the LLM completion just

:PrompterGenerate

The status line report some statistics:

Latency: 1480ms (1.5s) Tokens: 228 (prompt: 167 completion: 61) Throughput: 154 Words: 28 Chars: 176, Lines: 7

Explanation of values in the status line report:

          ┌─ latency in milliseconds and seconds
          │
          │                     ┌───────────────────────────────── total nr. of tokens
          │                     │
          │                     │            ┌──────────────────── nr. of tokens in prompt
          │                     │            │
          │                     │            │               ┌──── nr. of tokens in completion
          │                     │            │               │
        ┌─┴───────────┐       ┌─┴─┐        ┌─┴─┐           ┌─┴┐             ┌───┐      ┌──┐      ┌───┐       ┌─┐
Latency:│1480ms (1.5s)│Tokens:│228│(prompt:│167│completion:│61│) Throughput:│154│Words:│28│Chars:│176│ Lines:│7│
        └─────────────┘       └───┘        └───┘           └──┘             └─┬─┘      └─┬┘      └─┬─┘       └┬┘
                                                                              │          │         │          │
                                                                              │          │         │          │
                                      Latency / Tokens ───────────────────────┘          │         │          │
                                                                                         │         │          │
                                                                 nr. of words ───────────┘         │          │
                                                                                                   │          │
                                                            nr. of characters ─────────────────────┘          │
                                                                                                              │
                                                                 nr. of lines ────────────────────────────────┘

The statistics reports these metrics:

Metric	Description	Example
Latency	Bot in milliseconds and second approximation	`1480ms (1.5s)`
Tokens	Total tokens amount, prompt subtotal, and completion subtotal	`228`
Throughput	Completion Tokens / latency ratio (in seconds). See discussion about the concept of throughput here	`154`
Words	Number of words generated in the completion	`28`
Chars	Number of characters in the completion	`176`
Lines	Number of lines generated in the completion	`7`

By default the command is assigned to the function key F12. In such a way you can run the completion just pressing the single keystroke F12.

`:PrompterInfo`

Reports the current plugin version, the list of plugin commands, the current model settings.

:PrompterInfo

the command print these info:

Version:
prompter.vim, by giorgio.robino@gmail.com, version 0.2 (November 28, 2023)

Model:
Model: azure/gpt-35-turbo completion mode: chat temperature: 0.5 max_tokens: 1500

Commands:
PrompterGenerate   <F12>
PrompterInfo       <F10>
PrompterSetup      <F9>

🛠 Variables Settings

get and set completion background and foreground colors:
```
echo g:prompter_completion_ctermbg
echo g:prompter_completion_ctermfg

let g:prompter_completion_ctermbg = 3
let g:prompter_completion_ctermfg = 0
```
If you don't like the default highlight colors, you can replace ctermbg and ctermfg values using a subset of cterm/xterm 256 colors. To show all colors available you can use the command :HighlightColors part of my plugin: Highlight.
To modify the temperature value
```
let g:temperature = 0.2
```
To modify the max tokens value
```
let g:max_tokens = 2000
```
To modify the stop sequence(s)
```
let g:stop = ['x:', 'y:', 'z:']
```
Commands are associated to function keys with this default setting:
```
let g:prompter_setup_keystroke = '<F9>'
let g:prompter_info_keystroke = '<F10>'
let g:prompter_generate_keystroke =  '<F12>'
let g:prompter_regenerate_keystroke =  '<F8>'
```
Even if in vim you can assign a command to a key mapping of your preference, by example: map <F2> :PrompterGenerate<CR> and you can see what mapping for a particular key, e.g. F2, you can use the vim command: map <F12>, the suggested way to proceed is to modify one or more of the above mentioned variables and run :PrompterSetup again.

🛠 Other useful vim settings

Use (vim-included) spell checker

When you write a LLM prompt it's very very important to avoid typos! I many time experienced that LLM completion worst if you mistake just a verb.

Things go even worst if you are writing prompts in more than one languages. Personally I usually write conversational prompts in English, for some reasons described in my article Non-English Languages Prompt Engineering Trade-offs, but the target language of the chat prompt is my native language: Italian. All in all the prompt contains text in both English and Italian. In this case I run this small vimscript function:
```
function! Spell() 
   set spelllang=en_us,it
   setlocal spell
   echom "Spell check set for Italian and English languages"
endfunction  
com! SPELL call Spell()
```

Read all your previous completion statistics

messages

vim will show last completion statistics info. By example, if you just run 3 completions:

Latency: 961ms (1.0s) Tokens: 616 (prompt: 577 completion: 39) Throughput: 641 Words: 21 Chars: 134
Latency: 368ms (0.4s) Tokens: 648 (prompt: 642 completion: 6) Throughput: 1761 Words: 2 Chars: 15
Latency: 4227ms (4.2s) Tokens: 775 (prompt: 660 completion: 115) Throughput: 183 Words: 60 Chars: 377, Lines: 5

Enabling lines soft wrap

I usually work with a full-screen vim setting. That helps me to maximize my attention. Nevertheless having very long lines (after a PrompterGenerate) doesn't help the reading.
Unfortunately, in vim is not easy to configure a fixed column width soft wrap. See discussion. You can set set soft warp with following command:
```
set wrap linebreak nolist
```

💡 Dialogues as part of the text prompt

A technique I'm using to prototype dialog prompts, is to insert a dialog turns block as in the following example, where the dialog block terminates with the "stop sequence" (e.g. a:) triggering LLM to complete the assistant role:

TASK
You (a:) are a customer care assistant and you are assisting a user (u:).
...
...

DIALOG
a: Hello! How can I assist you today?
u: I want to open a report.
a: Fantastic! First, could you provide me with a detailed description of the issue you're experiencing?
u: The computer monitor won't turn on.
a: Thank you for the description. What is the name or model of the product or system you're having trouble with?
u: I can't read the brand. It's the company's PC monitor.
a: Thank you for the information. What is your preferred method of contact?
u: via email at g.angelotti@gmail.com
a: Thank you. Please confirm the provided email address: g.angelotti@gmail.com
u: that's correct!
a:

⚠ In the above case, to set the LLM stop waiting for user input, you could set the stop sequence as u: with command:

let g:stop = ['u:']

💡 Please note if you do not set the stop sequence as described above, the LLM will try to complete the entire conversation. This is in general not wanted because you want to write the sentence following u:. Nevertheless it's sometime useful to unset the g:stop just to see how the LLM imagine the conversation flow.

Other vim commands that could be useful:

Add a new line beginning with u: , just pressing the key F6:
```
map <F6> :normal ou: <CR> 
```
Add a new line beginning with a: , just pressing the key F7:
```
map <F7> :normal oa: <CR> 
```

Dev/Test Environment

Vim version: VIM - Vi IMproved 9.0
I developed on WLS (Window Linux Subsystem) on a Windows 10 operating system, using an Ubuntu 20.04 distribution.
I tested the plugin using (vim compiled for) Python 3.8.10
As LLM provider, I tested only my Azure OpenAI account (through LiteLLM library).
I didn't tested using an OpenAI account or other LLMs.

Change log

Version 0.1
First release. Only OpenAI/Azure OpenAI models are supported via OpenAI version 0.28 Python module.
Version 0.2
LLMs completions are done through LiteLLM, LLMs abstraction layer python package, allowing to use a multitude of different LLM providers.

Features to do in future releases

Support all LLM input parameters
So far prompter.vim support only temperature, max_tokens, stop arguments.
LiteLLM accepts and translates the OpenAI Chat Completion params across all providers.
Support template prompts
You are designing "template prompts" comprised of various parts that can be dynamically constructed at run-time. Consider, for instance, that you wish to prototype a "template prompt" containing placeholder variables, that are references to certain variables filled by other prompts or files, like so:
```
TASK
{some_task_description}

DATA
{some_yaml}

DIALOG
{dialog_history}
```
In the example above, when using web playgrounds, you function as a copy-paste intermediary. You are required to open four web tabs, execute text completions in each, and finally manually paste completions, substituting variables such as {some_data}, {dialog_history}. Additionally, you might need to load a file into a variable, like {some_yaml}.

The idea is to support template prompts editing allowing to replace on the fly (with a keystroke) the variable placeholders, with the content of other buffers/windows.
Asynchronous LLM completion
Currently, the LLM completion command PrompterGenerate is a synchronous command: the editor is blocked until the LLM API returns a completion text. It could be tedious for very complex and long prompts that require many many seconds to complete (e.g. >> 10). In this cases it could be better if the command could be asynchrous, allowing the developer to use the vim editor with the completion is on-going.
Streaming support
So far streaming completion is not take in consideration.

Known issues

in command Promptergenerate, some characters (to be defined) broke the highlight matchadd regexp. This imply that highlight doesn't work and/or a runtime error is generated. Nevertheless the generation is done. To be investigated.

Similar projects

vim-ai Very similar to prompter.vim, nevertheless it's focused on code completion allowing small prompts from the command line.
llm.nvim Just for neovim. Pretty similar to prompter.vim in the concept, but more oriented to code completions
llm.nvim just for neovim. It works with huggingface inference APis.
copilot.vim Neovim plugin for GitHub Copilot

👏 Acknowledgements

David Shapiro for his huge dissemination work on LLMs and generative AI. I have followed with enthusiasm especially his LLM prompt engineering live coding youtube videos!
Vivian De Smedt vim expert for his help solving an issue encountered when developing this plugin.
LiteLLM creators for having integrated some suggested features, as the Text Completion Format!

🙏 Status / How to contribute

This project is work-in-progress proof-of-concept alfa version!

I'm not a vimscript expert, so any contribute or suggestion is welcome. For any proposal and issue, please submit here on github issues for bugs, suggestions, etc. You can also contact me via email (giorgio.robino@gmail.com).

IF YOU LIKE THE PROJECT, PLEASE ⭐️STAR THIS REPOSITORY TO SHOW YOUR SUPPORT!

MIT LICENSE

Copyright (c) 2023 Giorgio Robino

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

top

Name		Name	Last commit message	Last commit date
Latest commit History 33 Commits
plugin		plugin
prompts		prompts
python		python
screens		screens
README.md		README.md

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

plugin

plugin

prompts

prompts

python

python

screens

screens

README.md

README.md

Repository files navigation

prompter.vim

How it works

⚠ A Tool for Prompt Engineers, Not Coders!

Features

Backstory

🙄 `text` or `chat` completion?

📦 Install

🔑 Environment Variables Setup

👊 Commands

`:PrompterSetup`

`:PrompterGenerate`

`:PrompterInfo`

🛠 Variables Settings

🛠 Other useful vim settings

💡 Dialogues as part of the text prompt

Dev/Test Environment

Change log

Features to do in future releases

Known issues

Similar projects

👏 Acknowledgements

🙏 Status / How to contribute

MIT LICENSE

About

Languages

solyarisoftware/prompter.vim

Folders and files

Latest commit

History

Repository files navigation

prompter.vim

How it works

⚠ A Tool for Prompt Engineers, Not Coders!

Features

Backstory

🙄 text or chat completion?

📦 Install

🔑 Environment Variables Setup

👊 Commands

:PrompterSetup

:PrompterGenerate

:PrompterInfo

🛠 Variables Settings

🛠 Other useful vim settings

💡 Dialogues as part of the text prompt

Dev/Test Environment

Change log

Features to do in future releases

Known issues

Similar projects

👏 Acknowledgements

🙏 Status / How to contribute

MIT LICENSE

About

Topics

Resources

Stars

Watchers

Forks

Languages

🙄 `text` or `chat` completion?

`:PrompterSetup`

`:PrompterGenerate`

`:PrompterInfo`