We are adding a new method for generating and storing embeddings in bulk, which relies on `Concurrent::Promises::Future`. Generating an embedding consists of three steps:
Prepare text
HTTP call to retrieve the vector
Save to DB.
Each one is independently executed on whatever thread the pool gives us.
We are bringing a custom thread pool instead of the global executor since we want control over how many threads we spawn to limit concurrency. We also avoid firing thousands of HTTP requests when working with large batches.
* PERF: Preload only gists when including summaries in topic list
* Add unique index on summaries and dedup existing records
* Make hot topics batch size setting hidden
* FIX: automatically bust cache for share ai assets
CDNs can be configured to strip query params in Discourse
hosting. This is generally safe, but in this case we had
no way of busting the cache using the path.
New design properly caches and properly breaks busts the
cache if asset changes so we don't need to worry about versions
* one day I will set up conditional lint on save :)
1. Keep source in a "details" block after rendered so it does
not overwhelm users
2. Ensure artifacts are never indexed by robots
3. Cache break our CSS that changed recently
It's important that artifacts are never given 'same origin' access to the forum domain, so that they cannot access cookies, or make authenticated HTTP requests. So even when visiting the URL directly, we need to wrap them in a sandboxed iframe.
This is a significant PR that introduces AI Artifacts functionality to the discourse-ai plugin along with several other improvements. Here are the key changes:
1. AI Artifacts System:
- Adds a new `AiArtifact` model and database migration
- Allows creation of web artifacts with HTML, CSS, and JavaScript content
- Introduces security settings (`strict`, `lax`, `disabled`) for controlling artifact execution
- Implements artifact rendering in iframes with sandbox protection
- New `CreateArtifact` tool for AI to generate interactive content
2. Tool System Improvements:
- Adds support for partial tool calls, allowing incremental updates during generation
- Better handling of tool call states and progress tracking
- Improved XML tool processing with CDATA support
- Fixes for tool parameter handling and duplicate invocations
3. LLM Provider Updates:
- Updates for Anthropic Claude models with correct token limits
- Adds support for native/XML tool modes in Gemini integration
- Adds new model configurations including Llama 3.1 models
- Improvements to streaming response handling
4. UI Enhancements:
- New artifact viewer component with expand/collapse functionality
- Security controls for artifact execution (click-to-run in strict mode)
- Improved dialog and response handling
- Better error management for tool execution
5. Security Improvements:
- Sandbox controls for artifact execution
- Public/private artifact sharing controls
- Security settings to control artifact behavior
- CSP and frame-options handling for artifacts
6. Technical Improvements:
- Better post streaming implementation
- Improved error handling in completions
- Better memory management for partial tool calls
- Enhanced testing coverage
7. Configuration:
- New site settings for artifact security
- Extended LLM model configurations
- Additional tool configuration options
This PR significantly enhances the plugin's capabilities for generating and displaying interactive content while maintaining security and providing flexible configuration options for administrators.
This re-implements tool support in DiscourseAi::Completions::Llm #generate
Previously tool support was always returned via XML and it would be the responsibility of the caller to parse XML
New implementation has the endpoints return ToolCall objects.
Additionally this simplifies the Llm endpoint interface and gives it more clarity. Llms must implement
decode, decode_chunk (for streaming)
It is the implementers responsibility to figure out how to decode chunks, base no longer implements. To make this easy we ship a flexible json decoder which is easy to wire up.
Also (new)
Better debugging for PMs, we now have a next / previous button to see all the Llm messages associated with a PM
Token accounting is fixed for vllm (we were not correctly counting tokens)
This change introduces a job to summarize topics and cache the results automatically. We provide a setting to control how many topics we'll backfill per hour and what the topic's minimum word count is to qualify.
We'll prioritize topics without summary over outdated ones.
Fixes encoding of params on LLM function calls.
Previously we would improperly return results if a function parameter returned an HTML tag.
Additionally adds some missing HTTP verbs to tool calls.
The custom field "discourse_ai_bypass_ai_reply" was added so
we can signal the post created hook to bypass replying even
if it thinks it should.
Otherwise there are cases where we double answer user questions
leading to much confusion.
This also slightly refactors code making the controller smaller
The new `/admin/plugins/discourse-ai/ai-personas/stream-reply.json` was added.
This endpoint streams data direct from a persona and can be used
to access a persona from remote systems leaving a paper trail in
PMs about the conversation that happened
This endpoint is only accessible to admins.
---------
Co-authored-by: Gabriel Grubba <70247653+Grubba27@users.noreply.github.com>
Co-authored-by: Keegan George <kgeorge13@gmail.com>
* FEATURE: Fast-track gist regeneration when a hot topic gets a new post
* DEV: Introduce an upsert-like summarize
* FIX: Only enqueue fast-track gist for hot hot hot topics
---------
Co-authored-by: Rafael Silva <xfalcox@gmail.com>
* FIX/REFACTOR: FoldContent revamp
We hit a snag with our hot topic gist strategy: the regex we used to split the content didn't work, so we cannot send the original post separately. This was important for letting the model focus on what's new in the topic.
The algorithm doesn’t give us full control over how prompts are written, and figuring out how to format the content isn't straightforward. This means we're having to use more complicated workarounds, like regex.
To tackle this, I'm suggesting we simplify the approach a bit. Let's focus on summarizing as much as we can upfront, then gradually add new content until there's nothing left to summarize.
Also, the "extend" part is mostly for models with small context windows, which shouldn't pose a problem 99% of the time with the content volume we're dealing with.
* Fix fold docs
* Use #shift instead of #pop to get the first elem, not the last
This changeset contains 4 fixes:
1. We were allowing running tests on unsaved tools,
this is problematic cause uploads are not yet associated or indexed
leading to confusing results. We now only show the test button when
tool is saved.
2. We were not properly scoping rag document fragements, this
meant that personas and ai tools could get results from other
unrelated tools, just to be filtered out later
3. index.search showed options as "optional" but implementation
required the second option
4. When testing tools searching through document fragments was
not working at all cause we did not properly load the tool
* FIX: Llm selector / forced tools / search tool
This fixes a few issues:
1. When search was not finding any semantic results we would break the tool
2. Gemin / Anthropic models did not implement forced tools previously despite it being an API option
3. Mechanics around displaying llm selector were not right. If you disabled LLM selector server side persona PM did not work correctly.
4. Disabling native tools for anthropic model moved out of a site setting. This deliberately does not migrate cause this feature is really rare to need now, people who had it set probably did not need it.
5. Updates anthropic model names to latest release
* linting
* fix a couple of tests I missed
* clean up conditional
A new feature_context json column was added to ai_api_audit_logs
This allows us to store rich json like context on any LLM request
made.
This new field now stores automation id and name.
Additionally allows llm_triage to specify maximum number of tokens
This means that you can limit the cost of llm triage by scanning only
first N tokens of a post.
This changeset:
1. Corrects some issues with "force_default_llm" not applying
2. Expands the LLM list page to show LLM usage
3. Clarifies better what "enabling a bot" on an llm means (you get it in the selector)
* Display gists in the hot topics list
* Adjust hot topics gist strategy and add a job to generate gists
* Replace setting with a configurable batch size
* Avoid loading summaries for other topic lists
* Tweak gist prompt to focus on latest posts in the context of the OP
* Remove serializer hack and rely on core change from discourse/discourse#29291
* Update lib/summarization/strategies/hot_topic_gists.rb
Co-authored-by: Rafael dos Santos Silva <xfalcox@gmail.com>
---------
Co-authored-by: Rafael dos Santos Silva <xfalcox@gmail.com>
Splits persona permissions so you can allow a persona on:
- chat dms
- personal messages
- topic mentions
- chat channels
(any combination is allowed)
Previously we did not have this flexibility.
Additionally, adds the ability to "tether" a language model to a persona so it will always be used by the persona. This allows people to use a cheaper language model for one group of people and more expensive one for other people
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
This PR updates the rate limits for AI helper so that image caption follows a specific rate limit of 20 requests per minute. This should help when uploading multiple files that need to be captioned. This PR also updates the UI so that it shows toast message with the extracted error message instead of having a blocking `popupAjaxError` error dialog.
---------
Co-authored-by: Rafael dos Santos Silva <xfalcox@gmail.com>
Co-authored-by: Penar Musaraj <pmusaraj@gmail.com>
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
Restructures LLM config page so it is far clearer.
Also corrects bugs around adding LLMs and having LLMs not editable post addition
---------
Co-authored-by: Sam Saffron <sam.saffron@gmail.com>
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.
Embedding search is rate limited due to potentially expensive
hyde operation (which require LLM access).
Embedding generally is very cheap compared to it. (usually 100x cheaper)
This raises the limit to 100 per minute for embedding searches,
while keeping the old 4 per minute for HyDE powered search.
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.