AI modules suddenly stopped working

RBoy · February 13, 2024, 5:23pm

I installed the Discourse AI plugin, setup Gemini and enabled a few modules, Summarization, Sentiment, Related Topics, AI Search and Toxicity.

At first it worked great, the Summarization button showed up and generated the content.

As soon as I enabled the Sentiment and Toxicity options, the Summarization stopped working, as in the button for Summarization disappeared from all topics. The sentiment tab on the dashboard shows no data to display.

When I look at the discourse logs, I see these errors:

Message

Job exception: ERROR:  must be owner of database discourse



Backtrace

rack-mini-profiler-3.3.0/lib/patches/db/pg.rb:110:in `exec' 
rack-mini-profiler-3.3.0/lib/patches/db/pg.rb:110:in `async_exec' 
mini_sql-1.5.0/lib/mini_sql/postgres/connection.rb:202:in `run' 
mini_sql-1.5.0/lib/mini_sql/active_record_postgres/connection.rb:38:in `block in run' 
mini_sql-1.5.0/lib/mini_sql/active_record_postgres/connection.rb:34:in `block in with_lock' 
activesupport-7.0.8/lib/active_support/concurrency/load_interlock_aware_monitor.rb:25:in `handle_interrupt' 
activesupport-7.0.8/lib/active_support/concurrency/load_interlock_aware_monitor.rb:25:in `block in synchronize' 
activesupport-7.0.8/lib/active_support/concurrency/load_interlock_aware_monitor.rb:21:in `handle_interrupt' 
activesupport-7.0.8/lib/active_support/concurrency/load_interlock_aware_monitor.rb:21:in `synchronize' 
mini_sql-1.5.0/lib/mini_sql/active_record_postgres/connection.rb:34:in `with_lock' 
mini_sql-1.5.0/lib/mini_sql/active_record_postgres/connection.rb:38:in `run' 
mini_sql-1.5.0/lib/mini_sql/postgres/connection.rb:181:in `exec' 
/var/www/discourse/plugins/discourse-ai/lib/embeddings/vector_representations/base.rb:154:in `create_index!' 
/var/www/discourse/plugins/discourse-ai/lib/embeddings/vector_representations/base.rb:71:in `block in consider_indexing' 
/var/www/discourse/plugins/discourse-ai/lib/embeddings/vector_representations/base.rb:49:in `each' 
/var/www/discourse/plugins/discourse-ai/lib/embeddings/vector_representations/base.rb:49:in `consider_indexing' 
/var/www/discourse/plugins/discourse-ai/app/jobs/scheduled/embeddings_backfill.rb:38:in `execute' 
/var/www/discourse/app/jobs/base.rb:297:in `block (2 levels) in perform' 
rails_multisite-5.0.0/lib/rails_multisite/connection_management.rb:82:in `with_connection'
/var/www/discourse/app/jobs/base.rb:284:in `block in perform' 
/var/www/discourse/app/jobs/base.rb:280:in `each' 
/var/www/discourse/app/jobs/base.rb:280:in `perform' 
/var/www/discourse/app/jobs/base.rb:351:in `perform' 
mini_scheduler-0.16.0/lib/mini_scheduler/manager.rb:122:in `process_queue' 
mini_scheduler-0.16.0/lib/mini_scheduler/manager.rb:70:in `worker_loop' 
mini_scheduler-0.16.0/lib/mini_scheduler/manager.rb:59:in `block (2 levels) in ensure_worker_threads'

Falco · February 28, 2024, 2:55pm

4 posts were split to a new topic: Discourse AI can’t set ivfflat.probes on self-hosted installs

Falco · February 15, 2024, 5:36pm

That means your database user can’t run

github.com

discourse/discourse-ai/blob/main/lib/embeddings/vector_representations/base.rb#L154


      
            DB.exec("COMMENT ON INDEX #{index_name} IS '#{Time.now.to_i}';")
            DB.exec("RESET work_mem;")
            DB.exec("RESET maintenance_work_mem;")
          
            database = DB.query_single("SELECT current_database();").first
          
            # This is a global setting, if we set it based on post count
            # we will be unable to use the index for topics
            # Hopefully https://github.com/pgvector/pgvector/issues/235 will make this better
            if table_name == topic_table_name
              DB.exec("ALTER DATABASE #{database} SET ivfflat.probes = #{probes};")
            end
          end
          
          def vector_from(text)
            raise NotImplementedError
          end
          
          def generate_representation_from(target, persist: true)
            text = @strategy.prepare_text_from(target, tokenizer, max_sequence_length - 2)
            return if text.blank?

on the Discourse database. I suggest checking your permissions.

RBoy · February 16, 2024, 2:00am

I’m sorry but I’m going to have a draw a blank on this one. I simply followed the instructions on this topic (and the other AI module topics), install the plugin, rebuild discourse and configure it from the admin page.

I didn’t see any step regarding configuring user permissions. Did I miss a step or could this be some exception scenario not handled in the install/build script?

RBoy · February 16, 2024, 9:25pm

How do I do this?
Shouldn’t this be automatically be set when I install the plugin and rebuild?

RBoy · February 20, 2024, 10:05pm

@Falco any thoughts on this?

Falco · February 28, 2024, 2:58pm

Looks like the database user on self-hosted installs lack the permissions to set this. A better way to set this is being tracked upstream, let’s wait for a fix there for a bit before trying to handle it

github.com/pgvector/pgvector

RFC: Setting search defaults as an index option

opened 03:04AM - 22 Aug 23 UTC

jkatz

## Use cases There are cases where a user wants to set a default search value… (e.g. `ivfflat.probes`, `hnsw.ef_search`) value on a per-index basis, for example: 1. Data is partitioned across multiple tables (integer-based partition key, not related to vector) with indexes on each table. This avoids setting the value on a per-query basis 2. User has a single index, but doesn't want to set the search values on a per-query basis ## Proposal The proposal is to add new index options on the index AMs: 1. `ivfflat`: `default_probes` (default: `NULL`) 1. `hnsw`: `default_ef_search` (default: `NULL`) PostgreSQL tracks when a [GUC is changed][1]. For this case, we can use the `PGC_S_SESSION` indicator to determine that GUC is changed in the database sessions and therefore we do not need to consult the index option. The proposal would then use this heuristic: ``` if default_probes/default_ef_search is NULL use value from `ivfflat.probes`/ `hnsw.ef_search` else if `ivfflat.probes`/`hnsw.ef_search` source is PGC_S_SESSION, use value of `ivfflat.probes`/`hnsw.ef_search` else use value from `default_probes`/`default_ef_search` ``` ## Alternatives With additional work, we could set "smarter" automatic defaults for `ivfflat.probes`/`hnsw.ef_search`, but folks may still want to set their own index-specific defaults based on the above use cases. Seeking feedback on this proposal. [1]: https://github.com/postgres/postgres/blob/master/src/include/utils/guc.h#L78-L123

RBoy · February 28, 2024, 3:29pm

Thanks for the update Falco. Could you please let me know what features or the impact of this and/or if it’s related to the log errors or some AI functionality that wont’ work; that I way I know what to expect or what won’t be working and can I need to temporarily disable those features to avoid other issues which may arise from this error.

EDIT: It looks like that number of probes being calculated based on the number of items in the database. Wondering if this could be set during a rebuild (so while the database may grow over time, each time discourse if upgraded, will optimize it until there’s a run time fix). Temporarily, at runtime check if there are SUPERUSER permissions before trying to alter it.

Falco · March 14, 2024, 8:20pm

This was fixed as part of the new AI quick search feature

RBoy · March 15, 2024, 12:33am

Fantastic. Upgrading now. Is there anything I need to do to kick start the previous process?

Falco · March 16, 2024, 11:00am

This topic was automatically closed after 38 hours. New replies are no longer allowed.

Topic		Replies	Views
Summarise with AI button... new? Feature ai , ai-summarize	26	1356	September 7, 2023
Discourse AI - Summarize Site Management how-to , ai , ai-summarize	33	5822	October 29, 2024
AI summarization not working after upgrading to 3.3.0 beta1 Support ai	11	575	April 19, 2024
Upgrade Error - rake db:migrate index on theme_field_id Installation two-container , ai	10	635	September 14, 2023
Upgrade failed with "duplicate key value violates unique constraint" error Installation	18	2331	March 30, 2020