On very large sites, the rare cache misses for Related Topics can take around 200ms, which affects our p99 metric on the topic page. In order to mitigate this impact, we now have several tools at our disposal.
First, one is to migrate the index embedding type from halfvec to bit and change the related topic query to leverage the new bit index by changing the search algorithm from inner product to Hamming distance. This will reduce our index sizes by 90%, severely reducing the impact of embeddings on our storage. By making the related query a bit smarter, we can have zero impact on recall by using the index to over-capture N*2 results, then re-ordering those N*2 using the full halfvec vectors and taking the top N. The expected impact is to go from 200ms to <20ms for cache misses and from a 2.5GB index to a 250MB index on a large site.
Another tool is migrating our index type from IVFFLAT to HNSW, which can increase the cache misses performance even further, eventually putting us in the under 5ms territory.
Co-authored-by: Roman Rizzi <roman@discourse.org>
This introduces another configuration that allows operators to
limit the amount of interactions with forced tool usage.
Forced tools are very handy in initial llm interactions, but as
conversation progresses they can hinder by slowing down stuff
and adding confusion.
This adds chain halting (ability to terminate llm chain in a tool)
and the ability to create uploads in a tool
Together this lets us integrate custom image generators into a
custom tool.
* FEATURE: allows forced LLM tool use
Sometimes we need to force LLMs to use tools, for example in RAG
like use cases we may want to force an unconditional search.
The new framework allows you backend to force tool usage.
Front end commit to follow
* UI for forcing tools now works, but it does not react right
* fix bugs
* fix tests, this is now ready for review
Previous to this change we could flag, but there was no way
to hide content and treat the flag as spam.
We had the option to hide topics, but this is not desirable for
a spam reply.
New option allows triage to hide a post if it is a reply, if the
post happens to be the first post on the topic, the topic will
be hidden.
This allows our users to add the Ollama provider and use it to serve our AI bot (completion/dialect).
In this PR, we introduce:
DiscourseAi::Completions::Dialects::Ollama which would help us translate by utilizing Completions::Endpoint::Ollama
Correct extract_completion_from and partials_from in Endpoints::Ollama
Also
Add tests for Endpoints::Ollama
Introduce ollama_model fabricator
This allows custom tools access to uploads and sophisticated searches using embedding.
It introduces:
- A shared front end for listing and uploading files (shared with personas)
- Backend implementation of index.search function within a custom tool.
Custom tools now may search through uploaded files
function invoke(params) {
return index.search(params.query)
}
This means that RAG implementers now may preload tools with knowledge and have high fidelity over
the search.
The search function support
specifying max results
specifying a subset of files to search (from uploads)
Also
- Improved documentation for tools (when creating a tool a preamble explains all the functionality)
- uploads were a bit finicky, fixed an edge case where the UI would not show them as updated
Caveats
- No streaming, by design
- No tool support (including no XML tools)
- No vision
Open AI will revamt the model and more of these features may
become available.
This solution is a bit hacky for now
Polymorphic RAG means that we will be able to access RAG fragments both from AiPersona and AiCustomTool
In turn this gives us support for richer RAG implementations.
Previously we waited 1 minute before automatically titling PMs
The new change introduces adding a title immediately after the the
llm replies
Prompt was also modified to include the LLM reply in title suggestion.
This helps situation like:
user: tell me a joke
llm: a very funy joke about horses
Then the title would be "A Funny Horse Joke"
Specs already covered some auto title logic, amended to also
catch the new message bus message we have been sending.
* FIX: we were never reindexing old content
Embedding backfill contains logic for searching for old content
change and then backfilling.
Unfortunately it was excluding all topics that had embedding
unconditionally, leading to no backfill ever happening.
This change adds a test and ensures we backfill.
* over select results, this ensures we will be more likely to find
ai results when filtered
This improves the site setting search so it performs a somewhat
fuzzy match.
Previously it did not handle seperators such as "space" and a
term such as "min_post_length" would not find "min_first_post_length"
A more liberal search algorithm makes it easier to the AI to
navigate settings.
* Minor fix, {{and parameter.enum parameter.enum.length}} is non
obviously broken.
If parameter.enum is a tracked array it will return the object
cause embers and helper implementation.
This corrects an issue where enum keeps on selecting itself by
mistake.
This allows callers of embedding based search to bypass hyde.
Hyde will expand the search term using an LLM, but if an LLM is
performing the search we can skip this expansion.
It also introduced some tests for the controller which we did not have
* FEATURE: LLM Triage support for systemless models.
This change adds support for OSS models without support for system messages. LlmTriage's system message field is no longer mandatory. We now send the post contents in a separate user message.
* Models using Ollama can also disable system prompts
New `ai_pm_summarization_allowed_groups` can be used to allow
visibility of the summarization feature on PMs.
This can be useful on forums where a lot of communication happens
inside PMs.
Creating a new model, either manually or from presets, doesn't initialize the `provider_params` object, meaning their custom params won't persist.
Additionally, this change adds some validations for Bedrock params, which are mandatory, and a clear message when a completion fails because we cannot build the URL.
* FIX: Add tool support to open ai compatible dialect and vllm
Automatic tools are in progress in vllm see: https://github.com/vllm-project/vllm/pull/5649
Even when they are supported, initial support will be uneven, only some models have native tool support
notably mistral which has some special tokens for tool support.
After the above PR lands in vllm we will still need to swap to
XML based tools on models without native tool support.
* fix specs
* DEV: Remove old code now that features rely on LlmModels.
* Hide old settings and migrate persona llm overrides
* Remove shadowing special URL + seeding code. Use srv:// prefix instead.
Using RAG fragments can lead to considerably big system messages, which becomes problematic when models have a smaller context window.
Before this change, we only look at the rest of the conversation to make sure we don't surpass the limit, which could lead to two unwanted scenarios when having large system messages:
All other messages are excluded due to size.
The system message already exceeds the limit.
As a result, I'm putting a hard-limit of 60% of available tokens. We don't want to aggresively truncate because if rag fragments are included, the system message contains a lot of context to improve the model response, but we also want to make room for the recent messages in the conversation.
Using assistant role for system produces an error because
they expect alternating roles like user/assistant/user and so on.
Prompts cannot start with the assistant role.
This allows summary to use the new LLM models and migrates of API key based model selection
Claude 3.5 etc... all work now.
---------
Co-authored-by: Roman Rizzi <rizziromanalejandro@gmail.com>
* FIX: Use base64 encoded images in AI Image Caption via LLaVa
This fixed a regression introduced in #646 where we started sending
schemaless URLs for our LLaVa service, which doesn't handle it well.
Moving to base64 encoded images solves:
- The service needing to download images
Now the service running LLaVa doesn't need internet access
- Secure uploads compat
Every image is treated the same, less branching for secure uploads
- Image Size problems
Discourse is now responsible for ensure a max size for images
- Troublesome dev env
Previously to this commit you would need a dev env that was internet
acessible to use llava image captions
Introduces custom AI tools functionality.
1. Why it was added:
The PR adds the ability to create, manage, and use custom AI tools within the Discourse AI system. This feature allows for more flexibility and extensibility in the AI capabilities of the platform.
2. What it does:
- Introduces a new `AiTool` model for storing custom AI tools
- Adds CRUD (Create, Read, Update, Delete) operations for AI tools
- Implements a tool runner system for executing custom tool scripts
- Integrates custom tools with existing AI personas
- Provides a user interface for managing custom tools in the admin panel
3. Possible use cases:
- Creating custom tools for specific tasks or integrations (stock quotes, currency conversion etc...)
- Allowing administrators to add new functionalities to AI assistants without modifying core code
- Implementing domain-specific tools for particular communities or industries
4. Code structure:
The PR introduces several new files and modifies existing ones:
a. Models:
- `app/models/ai_tool.rb`: Defines the AiTool model
- `app/serializers/ai_custom_tool_serializer.rb`: Serializer for AI tools
b. Controllers:
- `app/controllers/discourse_ai/admin/ai_tools_controller.rb`: Handles CRUD operations for AI tools
c. Views and Components:
- New Ember.js components for tool management in the admin interface
- Updates to existing AI persona management components to support custom tools
d. Core functionality:
- `lib/ai_bot/tool_runner.rb`: Implements the custom tool execution system
- `lib/ai_bot/tools/custom.rb`: Defines the custom tool class
e. Routes and configurations:
- Updates to route configurations to include new AI tool management pages
f. Migrations:
- `db/migrate/20240618080148_create_ai_tools.rb`: Creates the ai_tools table
g. Tests:
- New test files for AI tool functionality and integration
The PR integrates the custom tools system with the existing AI persona framework, allowing personas to use both built-in and custom tools. It also includes safety measures such as timeouts and HTTP request limits to prevent misuse of custom tools.
Overall, this PR significantly enhances the flexibility and extensibility of the Discourse AI system by allowing administrators to create and manage custom AI tools tailored to their specific needs.
Co-authored-by: Martin Brennan <martin@discourse.org>
Previously, we stored request parameters like the OpenAI organization and Bedrock's access key and region as site settings. This change stores them in the `llm_models` table instead, letting us drop more settings while also becoming more flexible.
* FEATURE: LLM presets for model creation
Previous to this users needed to look up complicated settings
when setting up models.
This introduces and extensible preset system with Google/OpenAI/Anthropic
presets.
This will cover all the most common LLMs, we can always add more as
we go.
Additionally:
- Proper support for Anthropic Claude Sonnet 3.5
- Stop blurring api keys when navigating away - this made it very complex to reuse keys
We no longer support the "provider:model" format in the "ai_helper_model" and
"ai_embeddings_semantic_search_hyde_model" settings. We'll migrate existing
values and work with our new data-driven LLM configs from now on.
Previously read tool only had access to public topics, this allows
access to all topics user has access to, if admin opts for the option
Also
- Fixes VLLM migration
- Display which llms have bot enabled
* DRAFT: Create AI Bot users dynamically and support custom LlmModels
* Get user associated to llm_model
* Track enabled bots with attribute
* Don't store bot username. Minor touches to migrate default values in settings
* Handle scenario where vLLM uses a SRV record
* Made 3.5-turbo-16k the default version so we can remove hack
- Display filtered search correctly, so it is not confusing
- When XML stripping, if a chunk was `<` it would crash
- SQL Helper improved to be better aware of Data Explorer
This is a rather huge refactor with 1 new feature (tool details can
be suppressed)
Previously we use the name "Command" to describe "Tools", this unifies
all the internal language and simplifies the code.
We also amended the persona UI to use less DToggles which aligns
with our design guidelines.
Co-authored-by: Martin Brennan <martin@discourse.org>
Native tools do not work well on Opus.
Chain of Thought prompting means it consumes enormous amounts of
tokens and has poor latency.
This commit introduce and XML stripper to remove various chain of
thought XML islands from anthropic prompts when tools are involved.
This mean Opus native tools is now functions (albeit slowly)
From local testing XML just works better now.
Also fixes enum support in Anthropic native tools
Add native Cohere tool support
- Introduce CohereTools class for tool translation and result processing
- Update Command dialect to integrate with CohereTools
- Modify Cohere endpoint to support passing tools and processing tool calls
- Add spec for testing tool triggering with Cohere endpoint
1. New tool to easily find files (and default branch) in a Github repo
2. Improved read tool with clearer params and larger context
* limit can totally mess up the richness semantic search adds, so include the results unconditionally.
Previoulsy on GPT-4-vision was supported, change introduces support
for Google/Anthropic and new OpenAI models
Additionally this makes vision work properly in dev environments
cause we sent the encoded payload via prompt vs sending urls
This change allows us to delete custom models. It checks if there is no module using them.
It also fixes a bug where the after-create transition wasn't working. While this prevents a model from being saved multiple times, endpoint validations are still needed (will be added in a separate PR).:
This is similar to code interpreter by ChatGPT, except that it uses
JavaScript as the execution engine.
Safeguards were added to ensure memory is constrained and evaluation
times out.
* FEATURE: Set endpoint credentials directly from LlmModel.
Drop Llama2Tokenizer since we no longer use it.
* Allow http for custom LLMs
---------
Co-authored-by: Rafael Silva <xfalcox@gmail.com>
When lazy load categories is enabled, the list of categories does not
have to fetched from the "site.json" endpoint because it is already
returned by "search.json".
This commit reverts commits 5056502 and 3e54697 because iterating over
all pages of categories is not really necessary.
- Introduce new support for GPT4o (automation / bot / summary / helper)
- Properly account for token counts on OpenAI models
- Track feature that was used when generating AI completions
- Remove custom llm support for summarization as we need better interfaces to control registration and de-registration
There are still some limitations to which models we can support with the `LlmModel` class. This will enable support for Llama3 while we sort those out.
This PR introduces the concept of "LlmModel" as a new way to quickly add new LLM models without making any code changes. We are releasing this first version and will add incremental improvements, so expect changes.
The AI Bot can't fully take advantage of this feature as users are hard-coded. We'll fix this in a separate PR.s
This optional feature allows search to be performed in the context
of the user that executed it.
By default we do not allow this behavior cause it means llm gets
access to potentially secure data.
When the bot is @mentioned, we need to be a lot more careful
about constructing context otherwise bot gets ultra confused.
This changes multiple things:
1. We were omitting all thread first messages (fixed)
2. Include thread title (if available) in context
3. Construct context in a clearer way separating user request from data
Both endpoints provide OpenAI-compatible servers. The only difference is that Vllm doesn't support passing tools as a separate parameter. Even if the tool param is supported, it ultimately relies on the model's ability to handle native functions, which is not the case with the models we have today.
As a part of this change, we are dropping support for StableBeluga/Llama2 models. They don't have a chat_template, meaning the new API can translate them.
These changes let us remove some of our existing dialects and are a first step in our plan to support any LLM by defining them as data-driven concepts.
I rewrote the "translate" method to use a template method and extracted the tool support strategies into its classes to simplify the code.
Finally, these changes bring support for Ollama when running in dev mode. It only works with Mistral for now, but it will change soon..
* Well, it was quite a journey but now tools have "context" which
can be critical for the stuff they generate
This entire change was so Dall E and Artist generate images in the correct context
* FIX: improve error handling around image generation
- also corrects image markdown and clarifies code
* fix spec
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.
Roman Rizzi
Rafael dos Santos Silva
Hoa Nguyen
Sam
Mark VanLandingham
Keegan George
PangBo
Loïc Guitaut
Bianca Nenciu