Artificial Intelligence

Elephant AI integrates with the Responses API to provide artificial intelligence to and from all Elephant modules.

The On-site chat

Site administrators can decide which group of users is able to use the chat with the AI models. Once in the chat, the user will be able to ask questions, and receive answers in natural language, using the same language he uses to ask the questions.

Questions can be related to the information provided or to Elephant modules data. For example, Which projects use JakartaEE?, Which partners offer this service?, What's the purpose of this site?, and so on.

Elephant AI SoC

Elephant AI also allows calling the AI models with operation-oriented instructions, to provide a more generic answer in natural language or a specific format.

Requirements

An API Key from a valid implementation.

Elephant AI provides

A set of Personalities, to provide different behaviors (currently CHAT and CREATOR).
Access to Personalities based on user's role.
Specific Developer and System instructions to ensure a correct behavior and compatibility with user language.
Tools from all modules to instruct responses with real data.
AI Contexts, any Elephant web page can be used to provide knowledge to the AI.
A set of documents with basic usage. Admins can decide whether to use them, or write their own.
A complete implementation of GitHub Flavored Markdown (GFM).
An On-site chat with the AI models.

Collect data in natural language, or in a specific format.
Upload data into files accessible to the models.
Content from site pages (automatic using AI Contexts).
Restrict questions within the configured context.
Options to add the Ask AI to the menu user widget, and/or a sticky button.

Length of Knowledge paradigm

This feature is under heavy development.

Length of Knowledge aims to provide different content (answers) to different kind of users, mainly based on user's role, but also user self identity. This model also has a big impact on token expenses.

Tools

User internal ID and roles.
Vector metadata.
Information mapping.
Unassisted model updates.

Disclosure

Files and set of files, based on metadata mapping.
Access to Elephant Graph.

Benefits

LoK per user.
User awareness.
Subjective questions.

AI Contexts

Elephant adds new context functionality related to the Elephant AI project.

Requirements

An API Key from a valid implementation.

Selecting a context to feed AI

When editing a context, you can select it as a feed for the AI models. It's important to keep in mind that the Vector DBs will be fed with the HTML content transformed to Markdown.

Keep the content easy to understand for your intended target. Create captions to clarify different sections. You can upload images for the web, but only the text will be used to feed the AI models. Use bold for important parts of the text, or italic, or monospaced.

As it happens with specific Markdown content for the AI models, it's important to follow the best practices. The AI models will cross-reference the information you provide, unless instructed not to (but what's the fun then?). You should review the nomenclature to allow good cross-references.

Elephant does the HTML to Markdown conversion automatically, so you don't have to worry about Markdown symbols. Write context content in normal Elephant Wiki syntax.

An example of Elephant Wiki using best practices:

In some context

---++Projects

The projects ...

In another context

See **Projects** to ...

Configuration

Elephant AI can be configured to respond to live-data. This is the list of properties, some associated to an AI Personality:

Property	Description
AI.API.Implementation	Could be OpenAI, Anthropic and Gemini.
AI.API.APIKEY	The API Key from a valid implementation.
AI.Run.Help	If set, it's an address to your own description of what the AI provides.
AI.Run.Debug	Adds information about the steps the AI is using.
AI.`Personality`.Model	The model used by the AI Personality.
AI.`Personality`.VectorStoreIDs	Storage ID where your selected content will be uploaded.
AI.`Personality`.AccessRole	The role required to interact with the AI Personality.
AI.`Personality`.IconButton	If true, shows a fixed position icon.
AI.`Personality`.DossierRoots	Category IDs of the projects will be sent to the AI Personality.
AI.`Personality`.ContactsRole	Contacts with the given roles will be sent to the AI Personality.
AI.`Personality`.UserDocs	User documentation will be sent to the AI Personality.

The user documentation is a set of documents included in Elephant AI. Their main purpose is to assist about common user actions, whilst mentioning that site policies and user role may allow or forbid some of these actions.

Currently, two vector stores are allowed.

The first might be used by normal users, with no special roles.
The second includes more private files, like users and their profiles. Only users with premium networking.

This is a brief description of current personalities and how they are used:

`CHAT`	The main personality, the one you'll be using first. The Ask AI chat uses this configuration. The features of this personality are: Restricted to your content. Linked to the AI Contexts you selected. Provides live-data to the AI.
`CREATOR`	The `CREATOR` personality aims to generate content. The main goal is to be attached to content editors. Whether is available to admins and/or users is controlled by the `AccesRole`, same as for `CHAT`. Not restricted, may answer any question. Does not provide live-data.

External content

External content can be made available to Elephant AI. The external content will be fed to the assistant using specific instructions and files.

The title is meant as a way for navigation. Can be overwritten by instructions.

Files are fed to the assistant. Keep filenames simple, and use the right file extension.

By default, Elephant AI provides all the instructions in English. Using a single language for instructions is a good practice, and should be enforced.

	Description
Title	The title is meant for the user to navigate contents. Keep the name simple and meaninful.
Active	Active AI content will be fed to the assistant. The main use of `active` is to allow editing without publishing.
Restricted	When restricted, the content is only available to `Premium Networking` users and `admins`. Otherwise, the content is available to any user with permission to access the assistant.
Instructions	The instructions will tell the assistant how and when to use the information provided by the files. At this point it's important to follow some rules: The recommended format for instructions is Markdown. Start the instructions with a header. By default, it would be `# Title Query Policy`. Despite being free to change this header, it's highly recommended for you to keep it. Include the context in which the files should be inspected. Be imperative about it. E.g. Use `file` ONLY when the user asks for `context`, replacing file with the filename and context with the keywords that represent this content. Mention all the files you have uploaded, and the reason why it should be inspected.
Documents	The documents you upload will be fed to the assistant. You can use as many formats as you want. Have in mind, though, that the most common formats will work better. That is, using plain text, with strong markup for starting/ending each item in the file, may work better. Keep in mind: Use strong field names, E.g. `name=Item name`. Use always the same fields for each item. E.g. do not `name=Item name` in one item and `title=Item name` in another. Whenever possible, use a unique entity value at top of the list of fields. Fields with much verbosity are better kept at the end of the field list.