mirror of
https://github.com/danny-avila/LibreChat.git
synced 2025-12-17 17:00:15 +01:00
fdc5265f48
* Create mkdocs.yaml * Create mkdocs.yml * Update mkdocs.yml * Update mkdocs.yml * Update mkdocs.yml * Update mkdocs.yml * Update mkdocs.yml * Update README.md * Update coding_conventions.md * Update documentation_guidelines.md * Update testing.md * Update heroku.md * Update hetzner_ubuntu.md * Update google_search.md * Update introduction.md * Update make_your_own.md * Update stable_diffusion.md * Update wolfram.md * Update proxy.md * Update user_auth_system.md * Update bing_jailbreak_info.md * Update breaking_changes.md * Update multilingual_information.md * Update project_origin.md * Update tech_stack.md * Update apis_and_tokens.md * Update docker_install.md * Update linux_install.md * Update mac_install.md * Update windows_install.md * Update mkdocs.yml * Update mkdocs.yml * Update documentation_guidelines.md * Add files via upload * Create temp.txt * Add files via upload * Delete logo.png * Create index.md * Update mkdocs.yml * Update mkdocs.yml * Delete temp.txt * Update README.md * Update README.md --------- Co-authored-by: Danny Avila <110412045+danny-avila@users.noreply.github.com>
75 lines
5.1 KiB
Markdown
75 lines
5.1 KiB
Markdown
# Plugins Endpoint
|
|
|
|
|
|
|
|
The plugins endpoint opens the door to prompting LLMs in new ways other than traditional input/output prompting.
|
|
|
|
The first step is using chain-of-thought prompting for using plugins/tools in a fashion mimicing the official ChatGPT Plugins feature.
|
|
|
|
More than this, you can use this endpoint for changing your conversation settings mid-conversation. Unlike the official ChatGPT site and all other endpoints, you can switch models, presets, and settings mid-convo, even when you have no plugins selected. This is useful if you first want a creative response from GPT-4, and then a deterministic, lower cost response from GPT-3. Soon, you will be able to use PaLM2 and HuggingFace models, all in this endpoint in the same modular manner.
|
|
|
|
### Roadmap:
|
|
- More plugins and advanced plugin usage
|
|
- More LLMs to choose from for both Thinking and Completion Phases
|
|
- Alternative prompting methods such as Tree-of-Thought
|
|
|
|
## Using Plugins
|
|
|
|
The LLM process when using Plugins is illustrated below.
|
|
|
|
|
|
|
|
**When you open the settings with the Plugins endpoint selected, you will view the default settings for the Completion Phase.**
|
|
|
|
Clicking on **"Show Agent Settings"** will allow you to modify parameters for the thinking phase
|
|
|
|
|
|
|
|
---
|
|
|
|
|
|
|
|
- You can specify which plugins you would like to select from by installing/uninstalling them in the Plugin store
|
|
- See this guide on how to create your own plugins (WIP)
|
|
- ChatGPT Plugins, both community-made and official versions, will be available for use in a future update
|
|
|
|
### Notes
|
|
- Every additional plugin selected will increase your token usage as there are detailed instructions the LLM needs for each one
|
|
- For best use, be selective with plugins per message and narrow your requests as much as possible
|
|
- If you need help coming up with a good plugin prompt, ask the LLM for suggestions before using one!
|
|
- Chain-of-thought prompting (plugin use) will always be more expensive than regular input/output prompting, so be sure it meets your need.
|
|
- Currently, the cheapest use will be to use gpt-3.5 for both phases
|
|
- From my testing, the best "bang for your buck" will be to use gpt-3.5 for the thinking phase, and gpt-4 for completion.
|
|
- Adding to above, if you ask for a poem and an image at the same time, it may work, but both may suffer in quality
|
|
- Instead, ask for a poem first with creative settings
|
|
- Then, ask for a good prompt for Stable Diffusion based on the poem
|
|
- Finally, use the Stable Diffusion plugin by referencing the pre-generated prompt
|
|
- Presets are only available when no Plugins are selected as the final review of the thinking phase has a specific system message.
|
|
- ⚠️ The **Browser/Scraper, Serpapi, and Zapier NLA plugins** are official langchain integrations and don't work the best. Improvements to them will be made
|
|
|
|
### Plugins Setup Instructions
|
|
- **[Google Search](./google_search.md)**
|
|
- **[Stable Diffusion](./stable_diffusion.md)**
|
|
- **[Wolfram](./wolfram.md)**
|
|
- **DALL-E** - same setup as above, you just need an OpenAI key, and it's made distinct from your main API key to make Chats but it can be the same one
|
|
- **Zapier** - You need a Zapier account. Get your [API key from here](https://nla.zapier.com/credentials/) after you've made an account
|
|
- Create allowed actions - Follow step 3 in this [getting start guide](https://nla.zapier.com/start/) from Zapier
|
|
- ⚠️ NOTE: zapier is known to be finicky with certain actions. I found that writing email drafts is probably the best use of it
|
|
- there are improvements that can be made to override the official NLA integration and that is TBD
|
|
- **Browser/Scraper** - This is not to be confused with 'browsing' on chat.openai.com (which is technically a plugin suite or multiple plugins)
|
|
- This plugin uses OpenAI embeddings so an OpenAI key is necessary, similar to DALL-E, and it's made distinct from your main API key to make Chats but it can be the same one
|
|
- This plugin will simply scrape html, and will not work with dynamic Javascript pages as that would require a more involved solution
|
|
- A better solution for 'browsing' is planned but can't guarantuee when
|
|
- This plugin is best used in combination with google so it doesn't hallucinate webpages to visit
|
|
- **Serpapi** - an alternative to Google search but not as performant in my opinion
|
|
- You can get an API key here: https://serpapi.com/dashboard
|
|
- For free tier, you are limited to 100 queries/month
|
|
- With google, you are limited to 100/day for free, which is a better deal, and any after may cost you a few pennies
|
|
|
|
### Showcase
|
|
|
|
|
|
|
|
|
|
|
|
|