Discourse AI causando novos erros de SSL e Connection Reset by Peer

[oznyet]

Priority/Severity:
Recent repository changes make Discourse AI mostly unworkable with current OpenAI API.

Platform:

• Self-hosted, using standard standalone build
• Ubuntu 24.04 host VM, Docker containers
• OpenAI API
• Anthropic API

Description:

Discourse AI has been calling an external API - OpenAI with models and working well as of Feb 15 (last container rebuild). Today (Feb 21) I rebuilt the container and things are not working.

Here’s what I know:

As at Feb 15
OpenAI models configured and working well:

• LLM/Persona
  • GPT4 Omni
  • GPT4 Omni Mini
• Embeddings
  • text-embedding-ada-002

As at Feb 21

All OpenAI models have about 70-80% error rate for LLM calls with the error message “Connection Reset by Peer”. Some chats go through, some fail partway in between. Embedding calls fail with an Faraday::ConnectionFailed SSL error.

Additional OpenAI models fail:

• o1-mini and o1-preview fail to test/save the LLM with a code error (‘developer’ is not a valid role) because developer role is only valid for o1 and o3 models, not their -mini versions. Source code github.com/discourse/discourse-ai/…/chat_gpt.rb:61 needs to be updated to do an exact model name match not a starts_with match. For the else case on line 73, there is no longer a system user and needs to be updated to be simply user. As of today, o1-mini cannot use tools.

Have tried:

• Checked OpenAI platform limits and we are well under Rate Limits and OpenAI account is funded.
• Rebuilding container
• Deleting and recreating LLM personas and users
• Deleting and creating LLM models
• Creating new API token keys
• Ensuring SSL and certificates were updated inside container
• Logging into container and using bash & curl to call API (successfully)
• Logging into rails console RAILS_ENV=production bundle exec rails console and using a http object to call the OpenAI API (successfully)
• Anthropic API calls to claude-3.5-sonnet (successfully)

Reproducible steps:

Create new container build using the latest Discourse and add the Discourse AI plugin to plugins:

  ...
  after_code:
    - exec:
        cd: $home/plugins
        cmd:
          - git clone https://github.com/discourse/discourse-ai.git

Configure OpenAI LLM and Embeddings models with the following:

• GPT4 Omni, GPT4 Omni Mini
  • All default values, insert your api key
  • Tokens: 64000
  • Click “Run test”, wait for response, sometimes success, often “Internal Server Error”. When successful, trying to chat with the persona gives the Inference LLM Model stack trace
• text-embedding-ada-002, text-embedding-3-large
  • Saves successfully, generates error logs, multiple repeating every 5 mins

Internal Server Error Stack Trace

Internal Server Error stack trace

Message (2 copies reported)
Errno::ECONNRESET (Connection reset by peer)
app/controllers/application_controller.rb:427:in `block in with_resolved_locale'
app/controllers/application_controller.rb:427:in `with_resolved_locale'
lib/middleware/omniauth_bypass_middleware.rb:35:in `call'
lib/content_security_policy/middleware.rb:12:in `call'
lib/middleware/anonymous_cache.rb:409:in `call'
lib/middleware/csp_script_nonce_injector.rb:12:in `call'
config/initializers/008-rack-cors.rb:26:in `call'
config/initializers/100-quiet_logger.rb:20:in `call'
config/initializers/100-silence_logger.rb:29:in `call'
lib/middleware/enforce_hostname.rb:24:in `call'
lib/middleware/processing_request.rb:12:in `call'
lib/middleware/request_tracker.rb:385:in `call'
Backtrace
openssl (3.3.0) lib/openssl/buffering.rb:217:in `sysread_nonblock'
openssl (3.3.0) lib/openssl/buffering.rb:217:in `read_nonblock'
net-protocol (0.2.2) lib/net/protocol.rb:218:in `rbuf_fill'
net-protocol (0.2.2) lib/net/protocol.rb:199:in `readuntil'
net-protocol (0.2.2) lib/net/protocol.rb:209:in `readline'
net-http (0.6.0) lib/net/http/response.rb:625:in `read_chunked'
net-http (0.6.0) lib/net/http/response.rb:595:in `block in read_body_0'
net-http (0.6.0) lib/net/http/response.rb:570:in `inflater'
net-http (0.6.0) lib/net/http/response.rb:593:in `read_body_0'
net-http (0.6.0) lib/net/http/response.rb:363:in `read_body'
plugins/discourse-ai/lib/completions/endpoints/base.rb:374:in `non_streaming_response'
plugins/discourse-ai/lib/completions/endpoints/base.rb:160:in `block (2 levels) in perform_completion!'
net-http (0.6.0) lib/net/http.rb:2433:in `block in transport_request'
net-http (0.6.0) lib/net/http/response.rb:320:in `reading_body'
net-http (0.6.0) lib/net/http.rb:2430:in `transport_request'
net-http (0.6.0) lib/net/http.rb:2384:in `request'
rack-mini-profiler (3.3.1) lib/patches/net_patches.rb:19:in `block in request_with_mini_profiler' 
rack-mini-profiler (3.3.1) lib/mini_profiler/profiling_methods.rb:44:in `step' 
rack-mini-profiler (3.3.1) lib/patches/net_patches.rb:18:in `request_with_mini_profiler' 
(eval at /var/www/discourse/lib/method_profiler.rb:38):12:in `request'
plugins/discourse-ai/lib/completions/endpoints/base.rb:122:in `block in perform_completion!'
net-http (0.6.0) lib/net/http.rb:1632:in `start'
net-http (0.6.0) lib/net/http.rb:1070:in `start'
plugins/discourse-ai/lib/completions/endpoints/base.rb:105:in `perform_completion!'
plugins/discourse-ai/lib/completions/endpoints/open_ai.rb:44:in `perform_completion!'
plugins/discourse-ai/lib/completions/llm.rb:281:in `generate'
plugins/discourse-ai/lib/configuration/llm_validator.rb:36:in `run_test'
plugins/discourse-ai/app/controllers/discourse_ai/admin/ai_llms_controller.rb:128:in `test'
actionpack (7.2.2.1) lib/action_controller/metal/basic_implicit_render.rb:8:in `send_action'
actionpack (7.2.2.1) lib/abstract_controller/base.rb:226:in `process_action'
actionpack (7.2.2.1) lib/action_controller/metal/rendering.rb:193:in `process_action'
actionpack (7.2.2.1) lib/abstract_controller/callbacks.rb:261:in `block in process_action'
activesupport (7.2.2.1) lib/active_support/callbacks.rb:121:in `block in run_callbacks'
app/controllers/application_controller.rb:427:in `block in with_resolved_locale'
i18n (1.14.7) lib/i18n.rb:353:in `with_locale'
app/controllers/application_controller.rb:427:in `with_resolved_locale'
activesupport (7.2.2.1) lib/active_support/callbacks.rb:130:in `block in run_callbacks'
activesupport (7.2.2.1) lib/active_support/callbacks.rb:141:in `run_callbacks'
actionpack (7.2.2.1) lib/abstract_controller/callbacks.rb:260:in `process_action'
actionpack (7.2.2.1) lib/action_controller/metal/rescue.rb:27:in `process_action'
actionpack (7.2.2.1) lib/action_controller/metal/instrumentation.rb:77:in `block in process_action'
activesupport (7.2.2.1) lib/active_support/notifications.rb:210:in `block in instrument'
activesupport (7.2.2.1) lib/active_support/notifications/instrumenter.rb:58:in `instrument'
activesupport (7.2.2.1) lib/active_support/notifications.rb:210:in `instrument'
actionpack (7.2.2.1) lib/action_controller/metal/instrumentation.rb:76:in `process_action'
actionpack (7.2.2.1) lib/action_controller/metal/params_wrapper.rb:259:in `process_action'
activerecord (7.2.2.1) lib/active_record/railties/controller_runtime.rb:39:in `process_action'
actionpack (7.2.2.1) lib/abstract_controller/base.rb:163:in `process'
actionview (7.2.2.1) lib/action_view/rendering.rb:40:in `process'
rack-mini-profiler (3.3.1) lib/mini_profiler/profiling_methods.rb:115:in `block in profile_method' 
actionpack (7.2.2.1) lib/action_controller/metal.rb:252:in `dispatch'
actionpack (7.2.2.1) lib/action_controller/metal.rb:335:in `dispatch'
actionpack (7.2.2.1) lib/action_dispatch/routing/route_set.rb:67:in `dispatch'
actionpack (7.2.2.1) lib/action_dispatch/routing/route_set.rb:50:in `serve'
actionpack (7.2.2.1) lib/action_dispatch/routing/mapper.rb:32:in `block in <class:Constraints>'
actionpack (7.2.2.1) lib/action_dispatch/routing/mapper.rb:62:in `serve'
actionpack (7.2.2.1) lib/action_dispatch/journey/router.rb:53:in `block in serve'
actionpack (7.2.2.1) lib/action_dispatch/journey/router.rb:133:in `block in find_routes'
actionpack (7.2.2.1) lib/action_dispatch/journey/router.rb:126:in `each'
actionpack (7.2.2.1) lib/action_dispatch/journey/router.rb:126:in `find_routes'
actionpack (7.2.2.1) lib/action_dispatch/journey/router.rb:34:in `serve'
actionpack (7.2.2.1) lib/action_dispatch/routing/route_set.rb:896:in `call'
lib/middleware/omniauth_bypass_middleware.rb:35:in `call'
rack (2.2.11) lib/rack/tempfile_reaper.rb:15:in `call'
rack (2.2.11) lib/rack/conditional_get.rb:27:in `call'
rack (2.2.11) lib/rack/head.rb:12:in `call'
actionpack (7.2.2.1) lib/action_dispatch/http/permissions_policy.rb:38:in `call'
lib/content_security_policy/middleware.rb:12:in `call'
lib/middleware/anonymous_cache.rb:409:in `call'
lib/middleware/csp_script_nonce_injector.rb:12:in `call'
config/initializers/008-rack-cors.rb:26:in `call'
rack (2.2.11) lib/rack/session/abstract/id.rb:266:in `context'
rack (2.2.11) lib/rack/session/abstract/id.rb:260:in `call'
actionpack (7.2.2.1) lib/action_dispatch/middleware/cookies.rb:704:in `call'
actionpack (7.2.2.1) lib/action_dispatch/middleware/callbacks.rb:31:in `block in call'
activesupport (7.2.2.1) lib/active_support/callbacks.rb:101:in `run_callbacks'
actionpack (7.2.2.1) lib/action_dispatch/middleware/callbacks.rb:30:in `call'
actionpack (7.2.2.1) lib/action_dispatch/middleware/debug_exceptions.rb:31:in `call'
actionpack (7.2.2.1) lib/action_dispatch/middleware/show_exceptions.rb:32:in `call'
logster (2.20.1) lib/logster/middleware/reporter.rb:40:in `call'
railties (7.2.2.1) lib/rails/rack/logger.rb:41:in `call_app'
railties (7.2.2.1) lib/rails/rack/logger.rb:29:in `call'
config/initializers/100-quiet_logger.rb:20:in `call'
config/initializers/100-silence_logger.rb:29:in `call'
actionpack (7.2.2.1) lib/action_dispatch/middleware/request_id.rb:33:in `call'
lib/middleware/enforce_hostname.rb:24:in `call'
rack (2.2.11) lib/rack/method_override.rb:24:in `call'
actionpack (7.2.2.1) lib/action_dispatch/middleware/executor.rb:16:in `call'
rack (2.2.11) lib/rack/sendfile.rb:110:in `call'
plugins/discourse-prometheus/lib/middleware/metrics.rb:14:in `call'
rack-mini-profiler (3.3.1) lib/mini_profiler.rb:334:in `call'
lib/middleware/processing_request.rb:12:in `call'
message_bus (4.3.9) lib/message_bus/rack/middleware.rb:60:in `call'
lib/middleware/request_tracker.rb:385:in `call'
actionpack (7.2.2.1) lib/action_dispatch/middleware/remote_ip.rb:96:in `call'
railties (7.2.2.1) lib/rails/engine.rb:535:in `call'
railties (7.2.2.1) lib/rails/railtie.rb:226:in `public_send'
railties (7.2.2.1) lib/rails/railtie.rb:226:in `method_missing'
rack (2.2.11) lib/rack/urlmap.rb:74:in `block in call'
rack (2.2.11) lib/rack/urlmap.rb:58:in `each'
rack (2.2.11) lib/rack/urlmap.rb:58:in `call'
unicorn (6.1.0) lib/unicorn/http_server.rb:634:in `process_client'
unicorn (6.1.0) lib/unicorn/http_server.rb:739:in `worker_loop'
unicorn (6.1.0) lib/unicorn/http_server.rb:547:in `spawn_missing_workers'
unicorn (6.1.0) lib/unicorn/http_server.rb:143:in `start'
unicorn (6.1.0) bin/unicorn:128:in `<top (required)>'
vendor/bundle/ruby/3.3.0/bin/unicorn:25:in `load'
vendor/bundle/ruby/3.3.0/bin/unicorn:25:in `<main>'

Viewing the log errors are:

Embeddings Model

Error message in logs: (every 5 mins) Connection reset by peer (Faraday::ConnectionFailed)

application_version: 00907363d4b290df1c755df1a2494b95265e40b4

job: Jobs::EmbeddingsBackfill

Embeddings model error Stack trace

Embeddings model error Stack trace

Job exception: 5 errors
Connection reset by peer (Faraday::ConnectionFailed)
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/openssl-3.3.0/lib/openssl/buffering.rb:217:in `sysread_nonblock'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/openssl-3.3.0/lib/openssl/buffering.rb:217:in `read_nonblock'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/net-protocol-0.2.2/lib/net/protocol.rb:218:in `rbuf_fill'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/net-protocol-0.2.2/lib/net/protocol.rb:199:in `readuntil'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/net-protocol-0.2.2/lib/net/protocol.rb:209:in `readline'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/net-http-0.6.0/lib/net/http/response.rb:625:in `read_chunked'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/net-http-0.6.0/lib/net/http/response.rb:595:in `block in read_body_0'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/net-http-0.6.0/lib/net/http/response.rb:570:in `inflater'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/net-http-0.6.0/lib/net/http/response.rb:593:in `read_body_0'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/net-http-0.6.0/lib/net/http/response.rb:363:in `read_body'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/net-http-0.6.0/lib/net/http/response.rb:401:in `body'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/net-http-0.6.0/lib/net/http/response.rb:321:in `reading_body'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/net-http-0.6.0/lib/net/http.rb:2430:in `transport_request'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/net-http-0.6.0/lib/net/http.rb:2384:in `request'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/rack-mini-profiler-3.3.1/lib/patches/net_patches.rb:19:in `block in request_with_mini_profiler'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/rack-mini-profiler-3.3.1/lib/mini_profiler/profiling_methods.rb:50:in `step'
/var/www/discourse/vendor/bundle/ruby/3.3.0/gems/rack-mini-profiler-3.3.1/lib/patches/net_patches.rb:18:in `request_with_mini_profil...
Backtrace
concurrent-ruby-1.3.5/lib/concurrent-ruby/concurrent/promises.rb:1268:in `raise' 
concurrent-ruby-1.3.5/lib/concurrent-ruby/concurrent/promises.rb:1268:in `wait_until_resolved!' 
concurrent-ruby-1.3.5/lib/concurrent-ruby/concurrent/promises.rb:998:in `value!' 
/var/www/discourse/plugins/discourse-ai/lib/embeddings/vector.rb:50:in `gen_bulk_reprensentations' 
/var/www/discourse/plugins/discourse-ai/app/jobs/scheduled/embeddings_backfill.rb:134:in `block in populate_topic_embeddings' 
/var/www/discourse/plugins/discourse-ai/app/jobs/scheduled/embeddings_backfill.rb:133:in `each' 
/var/www/discourse/plugins/discourse-ai/app/jobs/scheduled/embeddings_backfill.rb:133:in `each_slice' 
/var/www/discourse/plugins/discourse-ai/app/jobs/scheduled/embeddings_backfill.rb:133:in `populate_topic_embeddings' 
/var/www/discourse/plugins/discourse-ai/app/jobs/scheduled/embeddings_backfill.rb:36:in `execute' 
/var/www/discourse/app/jobs/base.rb:316:in `block (2 levels) in perform' 
rails_multisite-6.1.0/lib/rails_multisite/connection_management/null_instance.rb:49:in `with_connection'
rails_multisite-6.1.0/lib/rails_multisite/connection_management.rb:21:in `with_connection'
/var/www/discourse/app/jobs/base.rb:303:in `block in perform' 
/var/www/discourse/app/jobs/base.rb:299:in `each' 
/var/www/discourse/app/jobs/base.rb:299:in `perform' 
/var/www/discourse/app/jobs/base.rb:379:in `perform' 
mini_scheduler-0.18.0/lib/mini_scheduler/manager.rb:137:in `process_queue' 
mini_scheduler-0.18.0/lib/mini_scheduler/manager.rb:77:in `worker_loop' 
mini_scheduler-0.18.0/lib/mini_scheduler/manager.rb:63:in `block (2 levels) in ensure_worker_threads' 

Inference LLM Model

Error message in logs: Job exception: Connection reset by peer

application_version: 00907363d4b290df1c755df1a2494b95265e40b4

job: Jobs::CreateAiReply

LLM model error Stack trace

LLM model error Stack trace

Message
Job exception: Connection reset by peer
Backtrace
openssl-3.3.0/lib/openssl/buffering.rb:217:in `sysread_nonblock' 
openssl-3.3.0/lib/openssl/buffering.rb:217:in `read_nonblock' 
net-protocol-0.2.2/lib/net/protocol.rb:218:in `rbuf_fill' 
net-protocol-0.2.2/lib/net/protocol.rb:199:in `readuntil' 
net-protocol-0.2.2/lib/net/protocol.rb:209:in `readline' 
net-http-0.6.0/lib/net/http/response.rb:625:in `read_chunked' 
net-http-0.6.0/lib/net/http/response.rb:595:in `block in read_body_0' 
net-http-0.6.0/lib/net/http/response.rb:570:in `inflater' 
net-http-0.6.0/lib/net/http/response.rb:593:in `read_body_0' 
net-http-0.6.0/lib/net/http/response.rb:363:in `read_body' 
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:374:in `non_streaming_response' 
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:160:in `block (2 levels) in perform_completion!' 
net-http-0.6.0/lib/net/http.rb:2433:in `block in transport_request' 
net-http-0.6.0/lib/net/http/response.rb:320:in `reading_body' 
net-http-0.6.0/lib/net/http.rb:2430:in `transport_request' 
net-http-0.6.0/lib/net/http.rb:2384:in `request' 
rack-mini-profiler-3.3.1/lib/patches/net_patches.rb:19:in `block in request_with_mini_profiler' 
rack-mini-profiler-3.3.1/lib/mini_profiler/profiling_methods.rb:50:in `step' 
rack-mini-profiler-3.3.1/lib/patches/net_patches.rb:18:in `request_with_mini_profiler' 
(eval at /var/www/discourse/lib/method_profiler.rb:38):5:in `request'
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:122:in `block in perform_completion!' 
net-http-0.6.0/lib/net/http.rb:1632:in `start' 
net-http-0.6.0/lib/net/http.rb:1070:in `start' 
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:105:in `perform_completion!' 
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/open_ai.rb:44:in `perform_completion!' 
/var/www/discourse/plugins/discourse-ai/lib/completions/llm.rb:281:in `generate' 
/var/www/discourse/plugins/discourse-ai/lib/ai_bot/bot.rb:65:in `get_updated_title' 
/var/www/discourse/plugins/discourse-ai/lib/ai_bot/playground.rb:252:in `title_playground' 
/var/www/discourse/plugins/discourse-ai/lib/ai_bot/playground.rb:561:in `ensure in reply_to' 
/var/www/discourse/plugins/discourse-ai/lib/ai_bot/playground.rb:561:in `reply_to' 
/var/www/discourse/plugins/discourse-ai/app/jobs/regular/create_ai_reply.rb:18:in `execute' 
/var/www/discourse/app/jobs/base.rb:316:in `block (2 levels) in perform' 
rails_multisite-6.1.0/lib/rails_multisite/connection_management/null_instance.rb:49:in `with_connection'
rails_multisite-6.1.0/lib/rails_multisite/connection_management.rb:21:in `with_connection'
/var/www/discourse/app/jobs/base.rb:303:in `block in perform' 
/var/www/discourse/app/jobs/base.rb:299:in `each' 
/var/www/discourse/app/jobs/base.rb:299:in `perform' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:202:in `execute_job' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:170:in `block (2 levels) in process' 
sidekiq-6.5.12/lib/sidekiq/middleware/chain.rb:177:in `block in invoke' 
/var/www/discourse/lib/sidekiq/pausable.rb:132:in `call' 
sidekiq-6.5.12/lib/sidekiq/middleware/chain.rb:179:in `block in invoke' 
sidekiq-6.5.12/lib/sidekiq/middleware/chain.rb:182:in `invoke' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:169:in `block in process' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:136:in `block (6 levels) in dispatch' 
sidekiq-6.5.12/lib/sidekiq/job_retry.rb:113:in `local' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:135:in `block (5 levels) in dispatch' 
sidekiq-6.5.12/lib/sidekiq.rb:44:in `block in <module:Sidekiq>' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:131:in `block (4 levels) in dispatch' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:263:in `stats' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:126:in `block (3 levels) in dispatch' 
sidekiq-6.5.12/lib/sidekiq/job_logger.rb:13:in `call' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:125:in `block (2 levels) in dispatch' 
sidekiq-6.5.12/lib/sidekiq/job_retry.rb:80:in `global' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:124:in `block in dispatch' 
sidekiq-6.5.12/lib/sidekiq/job_logger.rb:39:in `prepare' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:123:in `dispatch' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:168:in `process' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:78:in `process_one' 
sidekiq-6.5.12/lib/sidekiq/processor.rb:68:in `run' 
sidekiq-6.5.12/lib/sidekiq/component.rb:8:in `watchdog' 
sidekiq-6.5.12/lib/sidekiq/component.rb:17:in `block in safe_thread' 

Any ideas, suggestions, etc would be most welcome

[Falco]

@oznyet

Todos os modelos da OpenAI têm uma taxa de erro de cerca de 70-80% para chamadas LLM com a mensagem de erro “Connection Reset by Peer”. Alguns chats funcionam, outros falham no meio. Chamadas de embedding falham com um erro SSL Faraday::ConnectionFailed.

Fizemos mais de 2 mil requisições para a API da OpenAI neste site no último mês, e não temos erros como esse em nossos logs.

Entrei em outro site que hospedamos, que usa intensamente a OpenAI, com 80 mil requisições no último mês e eles têm zero erros Faraday, e apenas dois erros “Connection Reset by Peer” no período.

É possível que haja algo errado na rede do seu servidor? Tive Errno::ECONNRESET (Connection reset by peer) uma vez por causa de um driver de NIC defeituoso.

@oznyet

no1-mini e o1-preview falham ao testar/salvar o LLM com um erro de código (‘developer’ não é uma função válida) porque a função ‘developer’ é válida apenas para os modelos o1 e o3, não para suas versões -mini. O código fonte github.com/discourse/discourse-ai/…/chat_gpt.rb:61 precisa ser atualizado para fazer uma correspondência exata do nome do modelo, não uma correspondência starts_with. Para o caso else na linha 73, não existe mais um usuário system e precisa ser atualizado para ser simplesmente user. A partir de hoje, o1-mini não pode usar ferramentas.

Vou verificar isso.

[oznyet]

Como você escreveu, inicialmente pensei que fosse um problema na pilha de rede, no entanto, chamadas repetidas da API OpenAI do próprio contêiner funcionam bem.

Isso também é com uma compilação muito recente, commits de 21 de fevereiro.

Para provar isso (ao custo de consumir tokens), criei um script rápido para verificar a pilha de rede da OpenAI.

• Executa por 600 segundos (10 minutos)
• Faz uma chamada de conclusão de chat por segundo
• Altera o prompt para evitar cache

Execute dentro do contêiner ./launcher enter app, salve o script abaixo, torne-o executável com chmod +x test_openai.sh e, em seguida, chame-o com OPENAI_API_KEY=.... ./test_openai.sh

test_openai.sh

#!/bin/bash

# Duração para executar
DURATION_SECS=600

# Inicializar contadores
successful=0
unsuccessful=0
declare -A error_messages

# Função para calcular porcentagem
calc_percentage() {
    local total=$(($1 + $2))
    if [ $total -eq 0 ]; then
        echo "0.00"
    else
        echo "scale=2; ($2 * 100) / $total" | bc
    fi
}

# Função para imprimir estatísticas
print_stats() {
    local percent=$(calc_percentage $successful $unsuccessful)
    echo "-------------------"
    echo "Chamadas bem-sucedidas: $successful"
    echo "Chamadas com falha: $unsuccessful"
    echo "Taxa de falha: ${percent}%"
    echo "Mensagens de erro:"
    for error in "${!error_messages[@]}"; do
        echo "  - $error (${error_messages[$error]} vezes)"
    done
}

end_time=$((SECONDS + DURATION_SECS))

counter=1
while [ $SECONDS -lt $end_time ]; do
    # Fazer a chamada da API com timeout
    response=$(curl -s -w "\n%{http_code}" \
        -X POST \
        -H "Content-Type: application/json" \
        -H "Authorization: Bearer $OPENAI_API_KEY" \
        -d "{
            \"model\": \"gpt-4o-mini\",
            \"messages\": [{\"role\": \"user\", \"content\": \"Use this number to choose a one word response: $counter\"}]
        }" \
        --connect-timeout 5 \
        --max-time 10 \
        https://api.openai.com/v1/chat/completions 2>&1)

    # Obter a última linha (código de status) e o corpo da resposta
    http_code=$(echo "$response" | tail -n1)
    body=$(echo "$response" | sed '$d')

    # Verificar se a chamada foi bem-sucedida
    if [ "$http_code" = "200" ]; then
        ((successful++))
    else
        ((unsuccessful++))
        # Extrair mensagem de erro
        error_msg=$(echo "$body" | grep -o '"message":"[^"]*"' | cut -d'"' -f4)
        if [ -z "$error_msg" ]; then
            error_msg="Connection error: $body"
        fi
        # Incrementar contador de mensagem de erro
        ((error_messages["$error_msg"]++))
    fi

    # Imprimir estatísticas atuais
    print_stats

    ((counter++))
    
    # Esperar 1 segundo antes da próxima chamada
    sleep 1
done

Para o script de teste, minha taxa de falha foi inferior a 0,5%, o que é aceitável nessa escala.

Isso me diz que o problema está no software Discourse e não no contêiner ou na pilha de rede que o alimenta.

Se não foi corrigido com um commit recente, darei uma olhada mais a fundo.

[sam (Sam Saffron)]

Corrigi a regressão em torno de o1-mini e o1-preview aqui:

FIX: legacy reasoning models not working, missing provider params

No entanto, estou confuso sobre os problemas de SSL, pois não alteramos nada em nossa biblioteca subjacente aqui.

Talvez isso esteja relacionado ao streaming, tente desativar o streaming em seu LLM da OpenAI e veja se isso se resolve? Seu teste lá está usando gpt-4o-mini sem usar streaming.

[oznyet]

@sam

Corrigi a regressão em torno de o1-mini e o1-preview aqui:

FIX: legacy reasoning models not working, missing provider params by SamSaffron · Pull Request #1149 · discourse/discourse-ai · GitHub

Isso é incrível! Bom trabalho!

No processo de diagnóstico, encontrei outro bug - na página de configuração do LLM (/admin/plugins/discourse-ai/ai-llms/%/edit), selecionar qualquer uma das opções para “Desabilitar suporte nativo de ferramentas (usar ferramentas baseadas em XML) (opcional)” ou “Desabilitar conclusões de streaming (converter requisições de streaming para não streaming)” e clicar em Salvar exibe uma notificação temporária de “Sucesso!”, mas ao recarregar a página, ambas ou qualquer uma das opções estará desmarcada.

Os problemas de reset de conexão ainda persistem e ainda estou investigando, no entanto, parece ser uma combinação do código Ruby (FinalDestination / Resolução de DNS / Faraday) de manipulação de sockets, combinado com o container Debian 12 em uma VM Ubuntu 24.04.

Eu criei uma VM de teste Ubuntu 22.04 e não há problemas, todos os embeddings e inferência funcionam perfeitamente. Não vi um único reset.

Continuarei a trabalhar nisso, talvez esteja relacionado a uma nova forma como o Ubuntu 24.04 gerencia a pilha TCP com netplan.

[sam (Sam Saffron)]

Obrigado, resolvi isso hoje a persistência, você pode atualizar e tentar novamente

[oznyet]

Ok, uma pequena atualização - não conseguimos fazer a conexão direta da API OpenAI funcionar no intervalo de IP corporativo. O Cloudflare enviava pacotes RST cerca de 1ms após o TLS.

Então, configuramos um Cloudflare AI Gateway como uma substituição de URL para o endpoint da API OpenAI e ele funciona perfeitamente com a configuração do LLM.

Parece que o Cloudflare tem uma política de limite de taxa não documentada para intervalos de IP desconhecidos (ou seja, não Azure, AWS, GCP, etc.) que entra em vigor. O pool de 100 conexões para Embeddings atingiria esse limite.

Como observação, o Cloudflare tem um recurso Authenticated Gateway que adiciona um token de cabeçalho especial.

Da documentação deles:

curl https://gateway.ai.cloudflare.com/v1/{account_id}/{gateway_id}/openai/chat/completions \
  --header 'cf-aig-authorization: Bearer {CF_AIG_TOKEN}' \
  --header 'Authorization: Bearer OPENAI_TOKEN' \
  --header 'Content-Type: application/json' \
  --data '{"model": "gpt-4o" .......

Seria ótimo se houvesse um recurso para adicionar cabeçalhos por LLM na tela de configuração do LLM.

Dessa forma, poderíamos adicionar a chave e o valor cf-aig-authorization ao LLM para cada chamada que fizermos.

[sam (Sam Saffron)]

Esta é uma tarefa difícil, é muita interface do usuário para lidar com um caso de uso específico.

Alguma chance de você tentar o openrouter.ai, pois isso também pode resolver este problema?

Não sou categoricamente contra permitir cabeçalhos arbitrários, mas é uma configuração muito avançada. Talvez se estivesse por trás de uma configuração oculta do site, estaria tudo bem (configuração do site para habilitar a interface do usuário avançada).

Esta é uma contribuição que sua empresa poderia ajudar a impulsionar para o plugin de código aberto?

[oznyet]

Ainda não consegui obter aprovação para uma contribuição, mas continuaremos tentando. Obrigado pela ajuda até agora!