This week’s discussions on meta.discourse.org revolved around exploring new AI models, configuring AI-powered spam detection, managing AI personas and prompts, and leveraging semantic search capabilities. Users shared their experiences with various language models, discussed potential improvements, and sought guidance on optimizing AI features within the Discourse platform.
Interesting Topics
Don initiated a discussion on Trying out the new AI models, sharing their experiences with models like grok-2-1212 and Gemini Flash 2. Challenges with language detection, formatting issues, and search capabilities were highlighted.
Sam announced a new AI-powered Spam detection module in Discourse AI, providing better control over spam scanning. They shared a configuration guide and encouraged users to share their experiences.
MarcP sought assistance with the Semantic Search API, and Sam provided guidance on accessing pure embeddings search and using API keys to bypass rate limits.
Don requested a Topic AI Summary box closing feature, allowing users to close the summary box by clicking outside it, which awesomerobot acknowledged as a reasonable request.
roniw raised concerns about privacy and data access when using Discourse’s default LLM providers, to which Sam responded, assuring that Discourse LLMs are self-hosted and data is not shared with third parties.
Don initiated the discussion on Trying out the new AI models, sharing their experiences and seeking advice from MihirR and Sam on improving language detection and search capabilities.
roniw raised concerns about privacy and data access when using Discourse’s default LLM providers, and Sam provided assurances about Discourse’s self-hosting and data privacy practices.
This week on the Discourse meta forum, discussions centered around the AI-powered spam detection feature, the proposed “Solved Finder” tool, issues with AI summarization, and various bug reports and feedback related to AI capabilities. Users shared insights, reported bugs, and suggested improvements, fostering a collaborative environment for enhancing the AI integration within the Discourse platform.
Interesting Topics
AI powered Spam detection: Users discussed the rollout of the AI-powered spam detection feature, including confusion around enabling it on hosted accounts and issues with the toggle behavior. @sam provided clarification on the process, and @Saif mentioned an upcoming fix for the toggle issue.
Solved finder: @Don proposed adding a button for staff users to find potential solutions within a topic, similar to the AI summary feature. @sam suggested exploring custom prompts or a “Talk to my topic” option to address this need.
AI Summary Backfill is wasting tokens summarizing PMs: @markschmucker raised a concern about the AI Summary Backfill process summarizing private messages (PMs), which could be inefficient. @Falco acknowledged the issue and mentioned the possibility of introducing a setting to skip PMs, similar to the existing setting for embeddings.
Can’t change ai_embeddings_model: @Overgrow reported an issue with changing the ai_embeddings_model setting and enabling embeddings via the admin panel. @sam provided a link to a pull request that aimed to fix the problem.
How Does Regenerating Summaries Work?: @markschmucker shared their experience with regenerating summaries using a different LLM and the challenges faced when the backfill process stopped after reaching the specified age limit, despite some topics failing due to rate limits.
Feedback on AI Summarisation: @markschmucker sought clarification on addressing invalid summaries, referring to a previous comment by @sam about potential fixes through refactoring or adjusting personas.
@MachineScholar shared their perspective on the dilemma of summarizing PMs, suggesting that the decision should be based on the community type and the potential value of generating embeddings for PMs.
@markschmucker shared their experience with regenerating summaries using a different LLM and the challenges faced when the backfill process stopped after reaching the specified age limit, despite some topics failing due to rate limits.
The past week on meta.discourse.org saw a diverse range of discussions around Discourse’s AI features, with users sharing their experiences, providing feedback, and seeking support. The community explored the integration of Discord and Discourse, the potential of AI-powered Q&A bots, and the capabilities of AI-powered spam detection. Additionally, there were conversations around the new Smart Dates feature in Helper, the AI Persona Editor, and the availability of AI features on Standard and Business plans.
Interesting Topics
Do you use Discord and Discourse? sparked a discussion about the simultaneous use of Discord and Discourse, their respective strengths, and the potential for an AI-powered Q&A Discord bot that leverages Discourse as a knowledge base.
Write out smarter dates with AI introduced the new Smart Dates feature in Helper, which converts human-typed times and dates into Discourse-compatible, timezone-friendly formats.
AI powered Spam detection highlighted the efficiency of Discourse’s AI-powered spam detection system, with users reporting positive experiences in identifying and hiding spam posts.
Discourse AI - AI triage explored the AI triage feature, which automates the categorization of topics using AI models, leading to discussions around the scalability and configuration of this feature.
Proofreading is too creative highlighted an instance where the AI proofreading feature in Helper made overly creative suggestions, prompting a discussion on the appropriate level of creativity for such features.
Feature requests for AI-Summarization gathered requests for improvements to the AI Summarization feature, including prioritizing the summarization of recently updated topics and the ability to skip summarizing personal messages.
How to hide bot completely, for everyone addressed a user’s request to hide the Discourse AI bot from their community, leading to a discussion on the limitations and potential solutions.
Discourse AI - Helper continued the ongoing discussion around the Helper feature, with users providing feedback and seeking clarification on its functionality.
This week on the meta.discourse.org forum, discussions revolved around various aspects of the Discourse AI plugin, including new features, improvements, and user queries. The AI-powered spam detection tool garnered attention, with users seeking clarification on its configuration and usage. Additionally, the “Ask Discourse” feature, which allows users to interact with an AI persona for documentation assistance, was introduced. Other topics included sortable sentiment tables, handling detailed content with AI, regenerating topic summaries, and managing AI usage quotas.
Interesting Topics
@per1234 highlighted an issue where the default LLMs were not selectable on the ‘Spam’ tab, despite being configured. @sam acknowledged the problem and provided a solution.
@Saif introduced the Ask Discourse feature, which allows users to interact with an AI persona to search through documentation and answer common questions related to Discourse.
@noahl requested the ability to sort sentiment tables by highest to lowest numbers, making it easier to identify and address common sentiments.
@c12gene raised a concern about the AI bot’s inability to read and process detailed content encapsulated in [summary] tags, prompting discussions on prompt engineering and tool usage.
@awesomerobot provided an update on the issue of the AI summary box closing, with a link to the relevant pull request.
@EricGT shared insights on LLM prompt evaluations, emphasizing the importance of understanding prompt effectiveness for LLM development.
@sam announced an upcoming feature for hard caps on user group AI usage for AI bots and AI Helper, allowing admins to define limits on total tokens and requests per group.
@per1234 reported a bug where posts and accounts were not always restored when a flag from the Discourse AI spam detection was rejected.
@Moin sought guidance on hiding the bot completely from all users, prompting a discussion on the visibility of the bot in search results.
Activity
@Saif introduced the Ask Discourse feature and provided details on its functionality and accessibility in the initial post. @Ed_Sshared their experience with the resource and reported a broken link, which @Saif later updated.
@EricGTshared insights on the importance of LLM prompt evaluations for effective prompt engineering.
@per1234reported a bug where posts and accounts were not always restored when a flag from the Discourse AI spam detection was rejected.
@Moinsought guidance on hiding the bot completely from all users, prompting a discussion with @Saif on the visibility of the bot in search results.
@BrianCinquired about the possibility of allowing users to upload documents for the AI bot to read and process, and @Saifclarified that this use case is currently not supported.
@MarcPasked about updating summaries for single post topics that are frequently edited.
Thanks for reading, and I’ll see you again next week!
This week on the meta.discourse.org forum, discussions revolved around various AI-related topics, including a proposed moderation tool for formatting code using AI, the ongoing development of Discourse AI features, issues with AI summarization backfills, and the potential for uploading and discussing PDFs within the composer. Additionally, there were conversations about setting usage limits for AI, evaluating costs between different AI providers, handling AI-generated spam, and exploring the capabilities of AI bots and custom tools.
Interesting Topics
@merefield raised an idea for a moderation tool that would allow trusted users to format code blocks using AI, potentially improving readability and assisting new users who struggle with proper code formatting.
Discussions continued on the Discourse AI plugin, with @sam expressing interest in allowing users to upload large files and ask questions about the content using a persona-based approach.
@markschmucker encountered an issue where the AI summarization backfill process kept regenerating summaries for the same topic, even after a valid summary was already present. This led to a fix by @Roman_Rizzi to make the job more resilient.
@BrianC proposed a feature request to allow users to upload PDFs or text files directly in the composer and have the AI process and respond to questions about the content.
There was a discussion around setting per-group token and usage limits for AI features, with @sam clarifying that quotas are defined per group and applied per user, rather than being shared among users.
There was a bug report regarding posts and accounts not being restored when flags from Discourse AI spam detection were rejected, which @sam addressed with a fix.
Discussions took place around self-hosting embeddings for DiscourseAI, with @sam mentioning ongoing work to restructure the embedding configuration and plans to support multi-model embeddings.
@David_Ghost inquired about the ability of AI Triage to perform searches and avoid topics with similar titles based on creation dates.
@David_Ghost inquired about the ability of AI Triage to perform searches and avoid topics with similar titles based on creation dates, with @sam indicating that such “agent-like” behaviors are being considered.
@c12gene reported an issue with the AI bot not being able to read summaries and detailed content, which was addressed by @MachineScholar through system prompt improvements.
@BrianC asked about tying token limits to subscriptions and allowing more expensive models to be used for a fee, with @sam confirming that different quotas can be set for different user groups.
