Add support for chat with AI personas
- Allow enabling chat for AI personas that have an associated user
- Add new setting `allow_chat` to AI persona to enable/disable chat
- When a message is created in a DM channel with an allowed AI persona user, schedule a reply job
- AI replies to chat messages using the persona's `max_context_posts` setting to determine context
- Store tool calls and custom prompts used to generate a chat reply on the `ChatMessageCustomPrompt` table
- Add tests for AI chat replies with tools and context
At the moment unlike posts we do not carry tool calls in the context.
No @mention support yet for ai personas in channels, this is future work
A recent change meant that llm instance got cached internally, repeat calls
to inference would cache data in Endpoint object leading model to
failures.
Both Gemini and Open AI expect a clean endpoint object cause they
set data.
This amends internals to make sure llm.generate will always operate
on clean objects
This commit introduces a new feature for AI Personas called the "Question Consolidator LLM". The purpose of the Question Consolidator is to consolidate a user's latest question into a self-contained, context-rich question before querying the vector database for relevant fragments. This helps improve the quality and relevance of the retrieved fragments.
Previous to this change we used the last 10 interactions, this is not ideal cause the RAG would "lock on" to an answer.
EG:
- User: how many cars are there in europe
- Model: detailed answer about cars in europe including the term car and vehicle many times
- User: Nice, what about trains are there in the US
In the above example "trains" and "US" becomes very low signal given there are pages and pages talking about cars and europe. This mean retrieval is sub optimal.
Instead, we pass the history to the "question consolidator", it would simply consolidate the question to "How many trains are there in the United States", which would make it fare easier for the vector db to find relevant content.
The llm used for question consolidator can often be less powerful than the model you are talking to, we recommend using lighter weight and fast models cause the task is very simple. This is configurable from the persona ui.
This PR also removes support for {uploads} placeholder, this is too complicated to get right and we want freedom to shift RAG implementation.
Key changes:
1. Added a new `question_consolidator_llm` column to the `ai_personas` table to store the LLM model used for question consolidation.
2. Implemented the `QuestionConsolidator` module which handles the logic for consolidating the user's latest question. It extracts the relevant user and model messages from the conversation history, truncates them if needed to fit within the token limit, and generates a consolidated question prompt.
3. Updated the `Persona` class to use the Question Consolidator LLM (if configured) when crafting the RAG fragments prompt. It passes the conversation context to the consolidator to generate a self-contained question.
4. Added UI elements in the AI Persona editor to allow selecting the Question Consolidator LLM. Also made some UI tweaks to conditionally show/hide certain options based on persona configuration.
5. Wrote unit tests for the QuestionConsolidator module and updated existing persona tests to cover the new functionality.
This feature enables AI Personas to better understand the context and intent behind a user's question by consolidating the conversation history into a single, focused question. This can lead to more relevant and accurate responses from the AI assistant.
This allows you to exclude trees of categories in a simple way
It also means you can no longer exclude "just the parent" but
this is a fair compromise.
- Adds support for sd3 and sd3 turbo models - this requires new endpoints
- Adds a hack to normalize arrays in the tool calls
- Removes some leftover code
- Adds support for aspect ratio as well so you can generate wide or tall images
For quite a few weeks now, some times, when running function calls
on Anthropic we would get a "stray" - "calls" line.
This has been enormously frustrating!
I have been unable to find the source of the bug so instead decoupled
the implementation and create a very clear "function call normalizer"
This new class is extensively tested and guards against the type of
edge cases we saw pre-normalizer.
This also simplifies the implementation of "endpoint" which no longer
needs to handle all this complex logic.
* FIX: various RAG edge cases
- Nicer text to describe RAG, avoids the word RAG
- Do not attempt to save persona when removing uploads and it is not created
- Remove old code that avoided touching rag params on create
* FIX: Missing pause button for persona users
* Feature: allow specific users to debug ai request / response chains
This can help users easily tune RAG and figure out what is going
on with requests.
* discourse helper so it does not explode
* fix test
* simplify implementation
* FEATURE: allow tuning of RAG generation
- change chunking to be token based vs char based (which is more accurate)
- allow control over overlap / tokens per chunk and conversation snippets inserted
- UI to control new settings
* improve ui a bit
* fix various reindex issues
* reduce concurrency
* try ultra low queue ... concurrency 1 is too slow.
Just having the word JSON can confuse models when we expect them
to deal solely in XML
Instead provide an example of how string arrays should be returned
Technically the tool framework supports int arrays and more, but
our current implementation only does string arrays.
Also tune the prompt construction not to give any tips about arrays
if none exist
This commit uses a new plugin modifier introduced in https://github.com/discourse/discourse/pull/26508
to mark all uploads as _not_ secure in shared PM AI conversations.
This is so images created by the AI bot (or uploaded by the user)
do not end up as broken URLs because of the security requirements
around them.
This relies on the UpdateTopicUploadSecurity job in core as well,
which is fired when an AI conversation is shared or deleted.
- Added Cohere Command models (Command, Command Light, Command R, Command R Plus) to the available model list
- Added a new site setting `ai_cohere_api_key` for configuring the Cohere API key
- Implemented a new `DiscourseAi::Completions::Endpoints::Cohere` class to handle interactions with the Cohere API, including:
- Translating request parameters to the Cohere API format
- Parsing Cohere API responses
- Supporting streaming and non-streaming completions
- Supporting "tools" which allow the model to call back to discourse to lookup additional information
- Implemented a new `DiscourseAi::Completions::Dialects::Command` class to translate between the generic Discourse AI prompt format and the Cohere Command format
- Added specs covering the new Cohere endpoint and dialect classes
- Updated `DiscourseAi::AiBot::Bot.guess_model` to map the new Cohere model to the appropriate bot user
In summary, this PR adds support for using the Cohere Command family of models with the Discourse AI plugin. It handles configuring API keys, making requests to the Cohere API, and translating between Discourse's generic prompt format and Cohere's specific format. Thorough test coverage was added for the new functionality.
BAAI/bge-m3 is an interesting model, that is multilingual and with a
context size of 8192. Even with a 16x larger context, it's only 4x slower
to compute it's embeddings on the worst case scenario.
Also includes a minor refactor of the rake task, including setting model
and concurrency levels when running the backfill task.
It used to fetch it from /site.json, but /categories.json is the more
appropriate one. This one also implements pagination, so we have to do
one request per page.
* FEATURE: Add metadata support for RAG
You may include non indexed metadata in the RAG document by using
[[metadata ....]]
This information is attached to all the text below and provided to
the retriever.
This allows for RAG to operate within a rich amount of contexts
without getting lost
Also:
- re-implemented chunking algorithm so it streams
- moved indexing to background low priority queue
* Baran gem no longer required.
* tokenizers is on 4.4 ... upgrade it ...
it is close in performance to GPT 4 at a fraction of the cost,
nice to add it to the mix.
Also improves a test case to simulate streaming, I am hunting for
the "calls" word that is jumping into function calls and can't quite
find it.
This PR lets you associate uploads to an AI persona, which we'll split and generate embeddings from. When building the system prompt to get a bot reply, we'll do a similarity search followed by a re-ranking (if available). This will let us find the most relevant fragments from the body of knowledge you associated with the persona, resulting in better, more informed responses.
For now, we'll only allow plain-text files, but this will change in the future.
Commits:
* FEATURE: RAG embeddings for the AI Bot
This first commit introduces a UI where admins can upload text files, which we'll store, split into fragments,
and generate embeddings of. In a next commit, we'll use those to give the bot additional information during
conversations.
* Basic asymmetric similarity search to provide guidance in system prompt
* Fix tests and lint
* Apply reranker to fragments
* Uploads filter, css adjustments and file validations
* Add placeholder for rag fragments
* Update annotations
This pull request makes several improvements and additions to the GitHub-related tools and personas in the `discourse-ai` repository:
1. It adds the `WebBrowser` tool to the `Researcher` persona, allowing the AI to visit web pages, retrieve HTML content, extract the main content, and convert it to plain text.
2. It updates the `GithubFileContent`, `GithubPullRequestDiff`, and `GithubSearchCode` tools to handle HTTP responses more robustly (introducing size limits).
3. It refactors the `send_http_request` method in the `Tool` class to follow redirects when specified, and to read the response body in chunks to avoid memory issues with large responses. (only for WebBrowser)
4. It updates the system prompt for the `Researcher` persona to provide more detailed guidance on when to use Google search vs web browsing, and how to optimize tool usage and reduce redundant requests.
5. It adds a new `web_browser_spec.rb` file with tests for the `WebBrowser` tool, covering various scenarios like handling different HTML structures and following redirects.
This commit adds the ability to enable vision for AI personas, allowing them to understand images that are posted in the conversation.
For personas with vision enabled, any images the user has posted will be resized to be within the configured max_pixels limit, base64 encoded and included in the prompt sent to the AI provider.
The persona editor allows enabling/disabling vision and has a dropdown to select the max supported image size (low, medium, high). Vision is disabled by default.
This initial vision support has been tested and implemented with Anthropic's claude-3 models which accept images in a special format as part of the prompt.
Other integrations will need to be updated to support images.
Several specs were added to test the new functionality at the persona, prompt building and API layers.
- Gemini is omitted, pending API support for Gemini 1.5. Current Gemini bot is not performing well, adding images is unlikely to make it perform any better.
- Open AI is omitted, vision support on GPT-4 it limited in that the API has no tool support when images are enabled so we would need to full back to a different prompting technique, something that would add lots of complexity
---------
Co-authored-by: Martin Brennan <martin@discourse.org>
We were only suppressing non mentions, ones that become spans.
@sam in the test was not resolving to a mention cause the user
did not exist.
depends on: https://github.com/discourse/discourse/pull/26253 for tests to pass.
- Stop replying as bot, when human replies to another human
- Reply as correct persona when replying directly to a persona
- Fix paper cut where suppressing notifications was not doing so
This PR consolidates the implements new Anthropic Messages interface for Bedrock Claude endpoints and adds support for the new Claude 3 models (haiku, opus, sonnet).
Key changes:
- Renamed `AnthropicMessages` and `Anthropic` endpoint classes into a single `Anthropic` class (ditto for ClaudeMessages -> Claude)
- Updated `AwsBedrock` endpoints to use the new `/messages` API format for all Claude models
- Added `claude-3-haiku`, `claude-3-opus` and `claude-3-sonnet` model support in both Anthropic and AWS Bedrock endpoints
- Updated specs for the new consolidated endpoints and Claude 3 model support
This refactor removes support for old non messages API which has been deprecated by anthropic
* FEATURE: allow suppression of notifications from report generation
Previously we needed to do this by hand, unfortunately this uses up
too many tokens and is very hard to discover.
New option means that we can trivially disable notifications without
needing any prompt engineering.
* URI.parse is safer, use it
* FIX: Handle unicode on tokenizer
Our fast track code broke when strings had characters who are longer in tokens than
in UTF-8.
Admins can set `DISCOURSE_AI_STRICT_TOKEN_COUNTING: true` in app.yml to ensure token counting is strict, even if slower.
Co-authored-by: wozulong <sidle.pax_0e@icloud.com>
* FIX: don't show share conversation incorrectly
- ai_persona_name can be null vs undefined leading to button showing up where it should not
- do not allow sharing of conversations where user is sending PMs to self
* remove erroneous code
* avoid query
This allows users to share a static page of an AI conversation with
the rest of the world.
By default this feature is disabled, it is enabled by turning on
ai_bot_allow_public_sharing via site settings
Precautions are taken when sharing
1. We make a carbonite copy
2. We minimize work generating page
3. We limit to 100 interactions
4. Many security checks - including disallowing if there is a mix
of users in the PM.
* Bonus commit, large PRs like this PR did not work with github tool
large objects would destroy context
Co-authored-by: Martin Brennan <martin@discourse.org>
Adds support for "name" on functions which can be used for tool calls
For function calls we need to keep track of id/name and previously
we only supported either
Also attempts to improve sql helper
Introduces a new AI Bot persona called 'GitHub Helper' which is specialized in assisting with GitHub-related tasks and questions. It includes the following key changes:
- Implements the GitHub Helper persona class with its system prompt and available tools
- Adds three new AI Bot tools for GitHub interactions:
- github_file_content: Retrieves content of files from a GitHub repository
- github_pull_request_diff: Retrieves the diff for a GitHub pull request
- github_search_code: Searches for code in a GitHub repository
- Updates the AI Bot dialects to support the new GitHub tools
- Implements multiple function calls for standard tool dialect
This provides new support for messages API from Claude.
It is required for latest model access.
Also corrects implementation of function calls.
* Fix message interleving
* fix broken spec
* add new models to automation
- FIX: only update system attributes when updating system persona
- FIX: update participant count by hand so bot messages show in inbox
Co-authored-by: Joffrey JAFFEUX <j.jaffeux@gmail.com>
* FIX: support multiple tool calls
Prior to this change we had a hard limit of 1 tool call per llm
round trip. This meant you could not google multiple things at
once or perform searches across two tools.
Also:
- Hint when Google stops working
- Log topic_id / post_id when performing completions
* Also track id for title
Previous to this fix if a tool call ever streamed a SPACE alone,
we would eat it and ignore it, breaking params
Also fixes some tests to ensure they are actually called :)
* DEV: improve internal design of ai persona and bug fix
- Fixes bug where OpenAI could not describe images
- Fixes bug where mentionable personas could not be mentioned unless overarching bot was enabled
- Improves internal design of playground and bot to allow better for non "bot" users
- Allow PMs directly to persona users (previously bot user would also have to be in PM)
- Simplify internal code
Co-authored-by: Martin Brennan <martin@discourse.org>
* FEATURE: AI helper support in non English languages
This attempts some prompt engineering to coerce AI helper to answer
in the appropriate language.
Note mileage will vary, in testing GPT-4 produces the best results
GPT-3.5 can return OKish results.
* Extend non english support for GPT-4V image caption
* Update db/fixtures/ai_helper/603_completion_prompts.rb
---------
Co-authored-by: Rafael Silva <xfalcox@gmail.com>
The Faraday adapter and `FinalDestionation::HTTP` will protect us from admin-initiated SSRF attacks when interacting with the external services powering this plugin features.:
Persona users are still bots, but we were not properly accounting
for it and share icon was not showing up.
This depends on a core change that adds .topic to transformed posts
Utilizes the check for secure upload permissions from core PR
https://github.com/discourse/discourse/pull/25758 and cleans up
controller codes and spec code to reuse existing code and better
reflect reality.
This PR adds a new feature where you can generate captions for images in the composer using AI.
---------
Co-authored-by: Rafael Silva <xfalcox@gmail.com>
This persona searches Discourse Meta for help with Discourse and
points users at relevant posts.
It is somewhat similar to using "Forum Helper" on meta, with the
notable difference that we can not lean on semantic search so using
some prompt engineering we try to keep it simple.
1. Personas are now optionally mentionable, meaning that you can mention them either from public topics or PMs
- Mentioning from PMs helps "switch" persona mid conversation, meaning if you want to look up sites setting you can invoke the site setting bot, or if you want to generate an image you can invoke dall e
- Mentioning outside of PMs allows you to inject a bot reply in a topic trivially
- We also add the support for max_context_posts this allow you to limit the amount of context you feed in, which can help control costs
2. Add support for a "random picker" tool that can be used to pick random numbers
3. Clean up routing ai_personas -> ai-personas
4. Add Max Context Posts so users can control how much history a persona can consume (this is important for mentionable personas)
Co-authored-by: Martin Brennan <martin@discourse.org>
* FIX: Better AI chat thread titles
- Fix quote removal when multi-line
- Use XML tags for better LLM output parsing
- Use stop_sequences for faster and less wasteful LLM calls
- Adds truncation as the last line of defense
* FEATURE: allow personas to supply top_p and temperature params
Code assistance generally are more focused at a lower temperature
This amends it so SQL Helper runs at 0.2 temperature vs the more
common default across LLMs of 1.0.
Reduced temperature leads to more focused, concise and predictable
answers for the SQL Helper
* fix tests
* This is not perfect, but far better than what we do today
Instead of fishing for
1. Draft sequence
2. Draft body
We skip (2), this means the composer "only" needs 1 http request to
open, we also want to eliminate (1) but it is a bit of a trickier
core change, may figure out how to pull it off (defer it to first draft save)
Value of bot drafts < value of opening bot conversations really fast
1. on failure we were queuing a job to generate embeddings, it had the wrong params. This is both fixed and covered in a test.
2. backfill embedding in the order of bumped_at, so newest content is embedded first, cover with a test
3. add a safeguard for hidden site setting that only allows batches of 50k in an embedding job run
Previously old embeddings were updated in a random order, this changes it so we update in a consistent order
* UX: Validations to Llm-backed features (except AI Bot)
This change is part of an ongoing effort to prevent enabling a broken feature due to lack of configuration. We also want to explicit which provider we are going to use. For example, Claude models are available through AWS Bedrock and Anthropic, but the configuration differs.
Validations are:
* You must choose a model before enabling the feature.
* You must turn off the feature before setting the model to blank.
* You must configure each model settings before being able to select it.
* Add provider name to summarization options
* vLLM can technically support same models as HF
* Check we can talk to the selected model
* Check for Bedrock instead of anthropic as a site could have both creds setup
* FEATURE: add support for new OpenAI embedding models
This adds support for just released text_embedding_3_small and large
Note, we have not yet implemented truncation support which is a
new API feature. (triggered using dimensions)
* Tiny side fix, recalc bots when ai is enabled or disabled
* FIX: downsample to 2000 items per vector which is a pgvector limitation
When you trim a prompt we never want to have a state where there
is a "tool" reply without a corresponding tool call, it makes no
sense
Also
- GPT-4-Turbo is 128k, fix that
- Claude was not preserving username in prompt
- We were throwing away unicode usernames instead of adding to
message
We're updating core to change TL based access settings to be group based. This requires some updates of tests to work correctly. (The existing test setup gives false positives.)
Account properly for function calls, don't stream through <details> blocks
- Rush cooked content back to client
- Wait longer (up to 60 seconds) before giving up on streaming
- Clean up message bus channels so we don't have leftover data
- Make ai streamer much more reusable and much easier to read
- If buffer grows quickly, rush update so you are not artificially waiting
- Refine prompt interface
- Fix lost system message when prompt gets long
* REFACTOR: Represent generic prompts with an Object.
* Adds a bit more validation for clarity
* Rewrite bot title prompt and fix quirk handling
---------
Co-authored-by: Sam Saffron <sam.saffron@gmail.com>
This PR introduces 3 things:
1. Fake bot that can be used on local so you can test LLMs, to enable on dev use:
SiteSetting.ai_bot_enabled_chat_bots = "fake"
2. More elegant smooth streaming of progress on LLM completion
This leans on JavaScript to buffer and trickle llm results through. It also amends it so the progress dot is much
more consistently rendered
3. It fixes the Claude dialect
Claude needs newlines **exactly** at the right spot, amended so it is happy
---------
Co-authored-by: Martin Brennan <martin@discourse.org>
Followup 2636efcd1b,
whenever ruby code was changed locally this would break
module loading, giving an "uninitialized constant
DiscourseAi::Embeddings::EntryPoint::SemanticRelated" error.
Previous to this change it was very hard to tell if completion was
stuck or not.
This introduces a "dot" that follows the completion and starts
flashing after 5 seconds.
* FIX: improve bot behavior
- Provide more information to Gemini context post function execution
- Use system prompts for Claude (fixes Dall E)
- Ensure Assistant is properly separated
- Teach Claude to return arrays in JSON vs XML
Also refactors tests so we do not copy tool preamble everywhere
* System msg is claude-2 only. fix typo
---------
Co-authored-by: Roman Rizzi <rizziromanalejandro@gmail.com>
We thought Azure's latest API version didn't have tool support yet, but I didn't understand it was complaining about a required field in the tool call message.
* FIX: don't include <details> in context
We need to be careful adding <details> into context of conversations
it can cause LLMs to hallucinate results
* Fix Gemini multi-turn ctx flattening
---------
Co-authored-by: Roman Rizzi <rizziromanalejandro@gmail.com>
DALL E command accepts an Array as a tool argument, this was not
parsed correctly by the invoker leading to errors generating
images with DALL E
Side quest ... don't use update! it calls validations and will now
fail due to email validation
It also corrects the syntax around tool support, which was wrong.
Gemini doesn't want us to include messages about previous tool invocations, so I had to shuffle around some code to send the response it generated from those invocations instead. For this, I created the "multi_turn" context, which bundles all the context involved in the interaction.
* DEV: AI bot migration to the Llm pattern.
We added tool and conversation context support to the Llm service in discourse-ai#366, meaning we met all the conditions to migrate this module.
This PR migrates to the new pattern, meaning adding a new bot now requires minimal effort as long as the service supports it. On top of this, we introduce the concept of a "Playground" to separate the PM-specific bits from the completion, allowing us to use the bot in other contexts like chat in the future. Commands are called tools, and we simplified all the placeholder logic to perform updates in a single place, making the flow more one-wayish.
* Followup fixes based on testing
* Cleanup unused inference code
* FIX: text-based tools could be in the middle of a sentence
* GPT-4-turbo support
* Use new LLM API
* FIX: AI helper not working correctly with mixtral
This PR introduces a new function on the generic llm called #generate
This will replace the implementation of completion!
#generate introduces a new way to pass temperature, max_tokens and stop_sequences
Then LLM implementers need to implement #normalize_model_params to
ensure the generic names match the LLM specific endpoint
This also adds temperature and stop_sequences to completion_prompts
this allows for much more robust completion prompts
* port everything over to #generate
* Fix translation
- On anthropic this no longer throws random "This is your translation:"
- On mixtral this actually works
* fix markdown table generation as well
* FEATURE: allow easy sharing of bot conversations
* Lean on new core API i
* Added system spec for copy functionality
* Update assets/javascripts/initializers/ai-bot-replies.js
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
* discourse later insted of setTimeout
* Update spec/system/ai_bot/share_spec.rb
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
* feedback from review
just check the whole payload
* remove uneeded code
* fix spec
---------
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
Previously endpoint/base would `+=` decoded_chunk to leftover
This could lead to cases where the leftover buffer had duplicate
previously processed data
Fix ensures we properly skip previously decoded data.
Introduce a Discourse Automation based periodical report. Depends on Discourse Automation.
Report works best with very large context language models such as GPT-4-Turbo and Claude 2.
- Introduces final_insts to generic llm format, for claude to work best it is better to guide the last assistant message (we should add this to other spots as well)
- Adds GPT-4 turbo support to generic llm interface
This PR adds tool support to available LLMs. We'll buffer tool invocations and return them instead of making users of this service parse the response.
It also adds support for conversation context in the generic prompt. It includes bot messages, user messages, and tool invocations, which we'll trim to make sure it doesn't exceed the prompt limit, then translate them to the correct dialect.
Finally, It adds some buffering when reading chunks to handle cases when streaming is extremely slow.:M
In https://github.com/discourse/discourse/pull/24740, `min_trust_to_create_topic` site setting was replaced by `create_topic_allowed_groups`. This PR replaces the former, deprecated one, with the latter.
This is somewhat experimental, but the context of likes/view/username
can help the llm find out what content is more important or even
common users that produce great content
This inflates the amount of tokens somewhat, but given it is all numbers
and search columns titles are only included once this is not severe
We were limiting to 20 results unconditionally cause we had to make
sure search always fit in an 8k context window.
Models such as GPT 3.5 Turbo (16k) and GPT 4 Turbo / Claude 2.1 (over 150k)
allow us to return a lot more results.
This means we have a much richer understanding cause context is far
larger.
This also allows a persona to tweak this number, in some cases admin
may want to be conservative and save on tokens by limiting results
This also tweaks the `limit` param which GPT-4 liked to set to tell
model only to use it when it needs to (and describes default behavior)
Keep in mind:
- GPT-4 is only going to be fully released next year - so this hardcodes preview model for now
- Fixes streaming bugs which became a big problem with GPT-4 turbo
- Adds Azure endpoing for turbo as well
Co-authored-by: Martin Brennan <martin@discourse.org>
Personas now support providing options for commands.
This PR introduces a single option "base_query" for the SearchCommand. When supplied all searches the persona will perform will also include the pre-supplied filter.
This can allow personas to search a subset of the forum (such as documentation)
This system is extensible we can add options to any command trivially.
c.f. de983796e1b66aa2ab039a4fb6e32cec8a65a098
There will soon be additional login_required checks
for Guardian, and the intent of many checks by automated
systems is better fulfilled by using BasicUser, which
simulates a logged in TL0 forum user, rather than an
anon user.
Previous to this change we relied on explicit loading for a files in Discourse AI.
This had a few downsides:
- Busywork whenever you add a file (an extra require relative)
- We were not keeping to conventions internally ... some places were OpenAI others are OpenAi
- Autoloader did not work which lead to lots of full application broken reloads when developing.
This moves all of DiscourseAI into a Zeitwerk compatible structure.
It also leaves some minimal amount of manual loading (automation - which is loading into an existing namespace that may or may not be there)
To avoid needing /lib/discourse_ai/... we mount a namespace thus we are able to keep /lib pointed at ::DiscourseAi
Various files were renamed to get around zeitwerk rules and minimize usage of custom inflections
Though we can get custom inflections to work it is not worth it, will require a Discourse core patch which means we create a hard dependency.
We must ensure we can isolate titles, and the models sometimes ignore the example we give them.
Additionally, anons can generate HyDE posts, so we need to check if user is nil when attempting to log requests.
* FEATURE: Azure OpenAI support for DALL*E 3
Previous to this there was no way to add an inference endpoint for
DALL*E on Azure cause it requires custom URLs
Also:
- On save, when editing a persona it would revert priority and enabled
- More forgiving parsing in command framework for array function calls
- By default generate HD images - they tend to be a bit better
- Improve DALL*E prompt which was getting very annoying and always echoing what it is about to do
- Add a bit of a sleep between retries on image generation
- Fix error handling in image_command
* FIX: no selected persona should pick first prioritized one
Previously we were looking at `.personaId` but there is only an
id attribute so it failed
* FEATURE: new DALL-E-3 persona
This persona generates images using DALL-E-3 API and is enabled
by default
Keep in mind that we are still waiting on seeds/gen_id so we can
not retain style consistently between turns.
This will change as soon as a new Open AI API provides the missing
parameters
Co-authored-by: Martin Brennan <martin@discourse.org>
Previous to this changeset we used a custom system for tools/command
support for Anthropic.
We defined commands by using !command as a signal to execute it
Following Anthropic Claude 2.1, there is an official supported syntax (beta)
for tools execution.
eg:
```
+ <function_calls>
+ <invoke>
+ <tool_name>image</tool_name>
+ <parameters>
+ <prompts>
+ [
+ "an oil painting",
+ "a cute fluffy orange",
+ "3 apple's",
+ "a cat"
+ ]
+ </prompts>
+ </parameters>
+ </invoke>
+ </function_calls>
```
This implements the spec per Anthropic, it should be stable enough
to also work on other LLMs.
Keep in mind that OpenAI is not impacted here at all, as it has its
own custom system for function calls.
Additionally:
- Fixes the title system prompt so it works with latest Anthropic
- Uses new spec for "system" messages by Anthropic
- Tweak forum helper persona to guide Anthropic a tiny be better
Overall results are pretty awesome and Anthropic Claude performs
really well now on Discourse
* Revert "FIX: We don't need to prepend anthropic. to bedrock models (#308)"
This reverts commit 8a01751991.
* FIX: Bedrock uses slightly different model names
* DEV: One LLM abstraction to rule them all
* REFACTOR: HyDE search uses new LLM abstraction
* REFACTOR: Summarization uses the LLM abstraction
* Updated documentation and made small fixes. Remove Bedrock claude-2 restriction
People tend to keep to 1 persona when working with the bot,
this adds local browser memory for the last persona you interacted
with so you do not need to select it over and over again.
This is per browser, not per user memory.
Also... clean up tests so they do not need to require stubs which
were breaking the build
---------
Co-authored-by: Martin Brennan <martin@discourse.org>
Introduces a UI to manage customizable personas (admin only feature)
Part of the change was some extensive internal refactoring:
- AIBot now has a persona set in the constructor, once set it never changes
- Command now takes in bot as a constructor param, so it has the correct persona and is not generating AIBot objects on the fly
- Added a .prettierignore file, due to the way ALE is configured in nvim it is a pre-req for prettier to work
- Adds a bunch of validations on the AIPersona model, system personas (artist/creative etc...) are all seeded. We now ensure
- name uniqueness, and only allow certain properties to be touched for system personas.
- (JS note) the client side design takes advantage of nested routes, the parent route for personas gets all the personas via this.store.findAll("ai-persona") then child routes simply reach into this model to find a particular persona.
- (JS note) data is sideloaded into the ai-persona model the meta property supplied from the controller, resultSetMeta
- This removes ai_bot_enabled_personas and ai_bot_enabled_chat_commands, both should be controlled from the UI on a per persona basis
- Fixes a long standing bug in token accounting ... we were doing to_json.length instead of to_json.to_s.length
- Amended it so {commands} are always inserted at the end unconditionally, no need to add it to the template of the system message as it just confuses things
- Adds a concept of required_commands to stock personas, these are commands that must be configured for this stock persona to show up.
- Refactored tests so we stop requiring inference_stubs, it was very confusing to need it, added to plugin.rb for now which at least is clearer
- Migrates the persona selector to gjs
---------
Co-authored-by: Joffrey JAFFEUX <j.jaffeux@gmail.com>
Co-authored-by: Martin Brennan <martin@discourse.org>
- New AiPersona model which can store custom personas
- Persona are restricted via group security
- They can contain custom system messages
- They can support a list of commands optionally
To avoid expensive DB calls in the serializer a Multisite friendly Hash was introduced (which can be expired on transaction commit)
This PR aims to clarify sentiment reports by replacing averages with a count of posts that have one of their values above a threshold (60), meaning we have some level of confidence they are, in fact, positive or negative.
Same thing happen with post emotions, with the difference that a post can have multiple values above it (30). Additionally, we dropped the "Neutral" axis.
We also reworded the tooltip next to each report title, and added an early return to signal we have no data available instead of displaying an empty chart.
This PR adds new reports for displaying information about post sentiments grouped by date and emotions group by TL.
Depends on discourse/discourse#24274
Function calling will start hallucinating if you reshape results.
Previously we were morphing from:
`{ prompts: ["prompt 1", "prompt 2"] }`
to
`{ prompts: { prompt: "prompt 1", seed: 222}, { ... `
This meant that over a few call sequences function_call starts hallucinating an incorrect shape.
This change grounds us even on GPT-3.5
This allows for 2 big features:
1. Artist can ship up to 4 prompts for image generation
2. Artist can regenerate images cause it is aware of seed
This allows for iteration on images maintaining visual style
Adds an AI Helper function when selecting text while viewing a topic.
---------
Co-authored-by: Keegan George <kgeorge13@gmail.com>
Co-authored-by: Roman Rizzi <roman@discourse.org>
Also fixes it so users without bot in header can send it messages.
Previous to this change we would seed all bots with database seeds.
This lead to lots of confusion for people who do not enable ai bot.
Instead:
1. We do not seed any bots **until** user enables the ai_bot_enabled setting
2. If it is disabled we will
a. If no messages were created by bot - delete it
b. Otherwise we will deactivate account
Under certain cases, for example:
```
there is this japanese band called kirimi, tell me more about them, try searching 3 times and at least 2 times in japanese before answering.
```
Results come back with blank snippets. This adds protection so this
is allowed and code does not simply blow up.
Per: https://platform.openai.com/docs/api-reference/authentication
There is an organization option which is useful for large orgs
> For users who belong to multiple organizations, you can pass a header to specify which organization is used for an API request. Usage from these API requests will count against the specified organization's subscription quota.
llm_triage supported claude 2 in triage, this implements it
OpenAI rate limits frequently, this introduces some exponential
backoff (3 attempts - 3 seconds, 9 and 27)
Also reduces temp of classifiers so they have consistent behavior
The new automation rule can be used to perform llm based classification and categorization of topics.
You specify a system prompt (which has %%POST%% as an input), if it returns a particular piece of text then we will apply rules such as tagging, hiding, replying or categorizing.
This can be used as a spam filter, a "oops you are in the wrong place" filter and so on.
Co-authored-by: Joffrey JAFFEUX <j.jaffeux@gmail.com>
This adds a new creative persona that has access to the underlying
model and no external integrations.
It allows people to use Claude/GPT models in a Discourse agnostic
way.
* FIX: properly truncate !command prompts
### What is going on here?
Previous to this change where a command was issued by the LLM it
could hallucinate a continuation eg:
```
This is what tags are
!tags
some nonsense here
```
This change introduces safeguards so `some nonsense here` does not
creep in to the prompt history, poisoning the llm results
This in effect grounds the llm a lot better and results in the llm
forgetting less about results.
The change only impacts Claude at the moment, but will also improve
stuff for llama 2 in future.
Also, this makes it significantly easier to test the bot framework
without an llm cause we avoid a whole bunch of complex stubbing
* blank is not a valid bot response, do not inject into prompt
We pass the text to the current LLM and ask them to generate a StableDifussion prompt.
We'll use that to generate 4 samples, temporarily creating uploads and returning their short URLs.
* FIX: Made bot more robust
This is a collection of small fixes
- Display "Searching for: ..." while searching instead of showing found 0 results.
- Only allow 5 commands in lang chain - 6 feels like too much
- On the 5th command stop informing the engine about functions, so it is forced to complete
- Add another 30 tokens of buffer and explain why
- Typo in command prompt
Co-authored-by: Alan Guo Xiang Tan <gxtan1990@gmail.com>
Note, we perform permission checks on tag list against anon
to ensure we do not disclose information about private tags
to the llm which could get extracted.
In specific scenarios (no special filters or limits) we will also
always include 5 semantic results (at least) with every query.
This effectively means that all very wide queries will always return
20 results, regardless of how complex they are.
Also:
FIX: embedding backfill rake task not working
We renamed internals, this corrects the implementation
* FEATURE: HyDE-powered semantic search.
It relies on the new outlet added on discourse/discourse#23390 to display semantic search results in an unobtrusive way.
We'll use a HyDE-backed approach for semantic search, which consists on generating an hypothetical document from a given keywords, which gets transformed into a vector and used in a asymmetric similarity topic search.
This PR also reorganizes the internals to have less moving parts, maintaining one hierarchy of DAOish classes for vector-related operations like transformations and querying.
Completions and vectors created by HyDE will remain cached on Redis for now, but we could later use Postgres instead.
* Missing translation and rate limiting
---------
Co-authored-by: Roman Rizzi <rizziromanalejandro@gmail.com>
The researcher persona has access to Google and can perform
various internet research tasks. At the moment it can not read
web pages, but that is under consideration
Previous to this change we relied on client side settings to
determine if an end user has access to the ai bot.
This meant that if a user was not aware they are a member of a
group (as it is with restricted visibility ones) they would not
see the bot button.
All checking has now moved to the server side, and tests were
added to cover.
This refactor changes it so we only include minimal data in the
system prompt which leaves us lots of tokens for specific searches
The new search command allows us to pull in settings on demand
Descriptions are include in short search results, and names only
in longer results
Also:
* In dev it is important to tell when calls are made to open ai
this adds a console log to increase awareness around token usage
* PERF: stop counting tokens so often
This changes it so we only count tokens once per response
Previously each time we heard back from open ai we would count
tokens, leading to uneeded delays
* bug fix, commands may reach in for tokenizer
* add logging to console for anthropic calls as well
* Update lib/shared/inference/openai_completions.rb
Co-authored-by: Martin Brennan <mjrbrennan@gmail.com>
Also adds ai_bot_enabled_personas so admins can tweak which stock
personas are enabled.
The new persona has a full listing of all site settings and is
able to get context for each setting.
This means you can ask it to search through settings for something
relevant.
Security wise there is no access to actual configuration of settings
just to the names / description and implementation.
Previously this was part of the forum helper persona however it
just clashes too much with other behaviors, isolating it makes
it far more powerful.
* sneaking this one in, user_emails is a non obvious table in our
structure.
usually one would assume users has emails so the clarifies a bit
better. plus it is a very common table to hit.
This splits out a bunch of code that used to live inside bots
into a dedicated concept called a Persona.
This allows us to start playing with multiple personas for the bot
Ships with:
artist - for making images
sql helper - for helping with data explorer
general - for everything and anything
Also includes a few fixes that make the generic LLM function implementation more robust
This command can be used to extract information about a discourse
site setting directly from source.
To operate it needs the rg binary in the container.
This fixes 2 big issues:
1. No matter how hard you try, grounding anthropic title prompt
is just too hard. This works around by only looking at the last
sentence it returns and treating as title
2. Non English locales would be stuck with "generic" title, this
ensures every bot message gets a title, using a custom field to
track
Also, slightly tunes some anthropic prompts.
Open AI support function calling, this has a very specific shape
that other LLMs have not quite adopted.
This simulates a command framework using system prompts on LLMs
that are not open AI.
Features include:
- Smart system prompt to steer the LLM
- Parameter validation (we ensure all the params are specified correctly)
This is being tested on Anthropic at the moment and intial results
are promising.
Azure requires a single HTTP endpoint per type of completion.
The settings: `ai_openai_gpt35_16k_url` and `ai_openai_gpt4_32k_url` can be
used now to configure the extra endpoints
This amends token limit which was off a bit due to function calls and fixes
a minor JS issue where we were not testing for a property
previously you would have to wait quite a while to see the prompt this implements
a very basic implementation of progress so you can see the API is working.
Also:
- Fix google progress.
- Handle the incredibly rare, zero results from google.
- Simplify command so it is less error prone
- replace invoke and attache results with a invoke
- ensure invoke can only ever be run once
- pass in all the information a command needs in constructor
- use new pattern throughout
- test invocation in isolation
- Attempt to hint reading is done by sending complete:true
- Do not include post_number in result unless it was sent in
- Rush visual feedback when a command is run (ensure we always revise)
- Include hyperlink in read command description
- Stop round tripping to GPT after image generation (speeds up images by a lot)
- Add a test for image command
This command is useful for reading a topics content. It allows us to perform
critical analysis or suggest answers.
Given 8k token limit in GPT-4 I hardcoded reading to 1500 tokens, but we can
follow up and allow larger windows on models that support more tokens.
On local testing even in this limited form this can be very useful.
* FIX: Google command was including full payload
Additionally there was no truncating happening meaning you could blow token
budget easily on a single search.
This made Google search mostly useless and it would mean that after using
Google we would revert to a clean slate which is very confusing.
* no need for nil there
The command framework had some confusing dispatching where it would dispatch
JSON blobs, this meant there was lots of parsing required in every command
The refactor handles transforming the args prior to dispatch which makes
consuming far simpler
This is also general prep to supporting some basic command framework in other
llms.
TopicQuery already provides a lot of safeguards and options for filtering topic, and enforcing permissions. It makes sense to rely on it as other plugins like discourse-assign do.
As a bonus, we now have access to the current_user while serializing these topics, so users will see things like unread posts count just like we do for the lists.
Claude 1 costs the same and is less good than Claude 2. Make use of Claude
2 in all spots ...
This also fixes streaming so it uses the far more efficient streaming protocol.
Single and multi-chunk summaries end using different prompts for the last summary. This change detects when the summarized content fits in a single chunk and uses a slightly different prompt, which leads to more consistent summary formats.
This PR also moves the chunk-splitting step to the `FoldContent` strategy as preparation for implementing streamed summaries.
* FEATURE: Embeddings to main db
This commit moves our embeddings store from an external configurable PostgreSQL
instance back into the main database. This is done to simplify the setup.
There is a migration that will try to import the external embeddings into
the main DB if it is configured and there are rows.
It removes support from embeddings models that aren't all_mpnet_base_v2 or OpenAI
text_embedding_ada_002. However it will now be easier to add new models.
It also now takes into account:
- topic title
- topic category
- topic tags
- replies (as much as the model allows)
We introduce an interface so we can eventually support multiple strategies
for handling long topics.
This PR severely damages the semantic search performance, but this is a
temporary until we can get adapt HyDE to make semantic search use the same
embeddings we have for semantic related with good performance.
Here we also have some ground work to add post level embeddings, but this
will be added in a future PR.
Please note that this PR will also block Discourse from booting / updating if
this plugin is installed and the pgvector extension isn't available on the
PostgreSQL instance Discourse uses.
* DEV: Better strategies for summarization
The strategy responsibility needs to be "Given a collection of texts, I know how to summarize them most efficiently, using the minimum amount of requests and maximizing token usage".
There are different token limits for each model, so it all boils down to two different strategies:
Fold all these texts into a single one, doing the summarization in chunks, and then build a summary from those.
Build it by combining texts in a single prompt, and truncate it according to your token limits.
While the latter is less than ideal, we need it for "bart-large-cnn-samsum" and "flan-t5-base-samsum", both with low limits. The rest will rely on folding.
* Expose summarized chunks to users
Reduce maximum replies to 2500 tokens and make them even for both GPT-3.5
and 4
Account for 400+ tokens in function definitions (this was unaccounted for)
* FEATURE: add ai_bot_enabled_chat commands and tune search
This allows admins to disable/enable GPT command integrations.
Also hones search results which were looping cause the result did not denote
the failure properly (it lost context)
* include more context for google command
include more context for time command
* type
The new site settings:
ai_openai_gpt35_url : distribution for GPT 16k
ai_openai_gpt4_url: distribution for GPT 4
ai_openai_embeddings_url: distribution for ada2
If untouched we will simply use OpenAI endpoints.
Azure requires 1 URL per model, OpenAI allows a single URL to serve multiple models. Hence the new settings.
```
prompt << build_message(bot_user.username, reply)
```
Would store a "cooked" prompt which is invalid, instead just store the raw
values which are later passed to build_message
Additionally:
1. Disable summary command which needs honing
2. Stop storing decorations (searched for X) in prompt which leads to straying
3. Ship username directly to model, avoiding "user: content" in prompts. This
was causing GPT to stray
Given latest GPT 3.5 16k which is both better steered and supports functions
we can now support rich bot integration.
Clunky system message based steering is removed and instead we use the
function framework provided by Open AI
* DEV: Remove the summarization feature
Instead, we'll register summarization implementations for OpenAI, Anthropic, and Discourse AI using the API defined in discourse/discourse#21813.
Core and chat will implement features on top of these implementations instead of this plugin extending them.
* Register instances that contain the model, requiring less site settings
Previous to this change we were chaining stuff too late and would execute
commands serially leading to very unexpected results
This corrects this and allows us to run stuff like:
> Search google 3/4 times on various permutations of
QUERY and answer this question.
We limit at 5 commands to ensure there are not pathological user cases
where you lean on the LLM to flood us with results.
For the time being smart commands only work consistently on GPT 4.
Avoid using any smart commands on the earlier models.
Additionally adds better error handling to Claude which sometimes streams
partial json and slightly tunes the search command.
blog.start_gpt_chat -> was on my blog
This also slightly tunes the search prompt to support filtering by oldest
and try a tiny bit harder to guide GPT 3.5 which is a bit of a losing battle
Co-authored-by: Krzysztof Kotlarek <kotlarek.krzysztof@gmail.com>
The rails_failover middleware will intercept all `PG::ConnectionBad` errors and put the cluster into readonly mode. It does not have any handling for multiple databases. Therefore, an issue with the embeddings database was taking the whole cluster into readonly.
This commit fixes the issue by rescuing `PG::Error` from all AI database accesses, and re-raises errors with a different class. It also adds a spec to ensure that an embeddings database outage does not affect the functionality of the topics/show route.
Co-authored-by: David Taylor <david@taylorhq.com>
Sam
Keegan George
Roman Rizzi
Rafael dos Santos Silva
Bianca Nenciu
Martin Brennan
Natalie Tay
Loïc Guitaut
Kris
Jarek Radosz
Gerhard Schlager
Ted Johansson
Jan Cernik
Krzysztof Kotlarek
David Taylor
David Battersby
Alan Guo Xiang Tan