Configurações para embeddings Hugging Face bge-large-en? -> bots RAG não responsivos!

[StevePlex (Steve King)]

Por favor, aconselhe.. quais são as melhores configurações para permitir que o modelo de incorporação bge-large-en funcione como o serviço de vetor padrão do Discourse AI?
Tenho uma instância bge-large-en em execução na AWS e sei que o meu Discourse AI está a comunicar com ela (ver teste abaixo), mas a incorporação não está a funcionar em geral (a incorporação OpenAI funciona bem).
RESUMO DO PROBLEMA: Bots RAG não respondem quando a incorporação está definida para HF bge-large-en
aqui está o modelo de incorporação da AWS:

aqui estão as configurações do Discourse AI:

aqui está um teste ‘Run test’ de um LLM personalizado do Discourse apenas para verificar a conectividade..

aqui estão os logs do bge-large-en no lado da AWS:

Muito obrigado!!

Aqui está o log de erro..

Job exception: can't quote Array

hostname ai-qa-ubuntu-s-1vcpu-2gb-amd-sfo3-01-app
process_id 1165935
application_version f9192835a7e4d2067c3d1844f43f9e7b69de39e7
current_db default
current_hostname ai-qa.net
job Jobs::CreateAiReply
problem_db default
time 7:22 pm
opts post_id 618
--- --- --- ---
--- ---
bot_user_id -1208
persona_id 5
current_site_id default


Backtrace

/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activerecord-7.0.8.1/lib/active_record/connection_adapters/abstract/quoting.rb:25:in `quote'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activerecord-7.0.8.1/lib/active_record/connection_adapters/postgresql/quoting.rb:69:in `quote'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activerecord-7.0.8.1/lib/active_record/connection_adapters/abstract/quoting.rb:51:in `quote_bound_value'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activerecord-7.0.8.1/lib/active_record/sanitization.rb:193:in `block in quote_bound_value'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activerecord-7.0.8.1/lib/active_record/sanitization.rb:193:in `map!'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activerecord-7.0.8.1/lib/active_record/sanitization.rb:193:in `quote_bound_value'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activerecord-7.0.8.1/lib/active_record/sanitization.rb:171:in `replace_bind_variable'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activerecord-7.0.8.1/lib/active_record/sanitization.rb:180:in `block in replace_named_bind_variables'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activerecord-7.0.8.1/lib/active_record/sanitization.rb:176:in `gsub'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activerecord-7.0.8.1/lib/active_record/sanitization.rb:176:in `replace_named_bind_variables'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activerecord-7.0.8.1/lib/active_record/sanitization.rb:128:in `sanitize_sql_array'
/var/www/discourse/lib/mini_sql_multisite_connection.rb:21:in `public_send'
/var/www/discourse/lib/mini_sql_multisite_connection.rb:21:in `encode'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/mini_sql-1.5.0/lib/mini_sql/connection.rb:64:in `to_sql'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/mini_sql-1.5.0/lib/mini_sql/postgres/connection.rb:202:in `run'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/mini_sql-1.5.0/lib/mini_sql/active_record_postgres/connection.rb:38:in `block in run'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/mini_sql-1.5.0/lib/mini_sql/active_record_postgres/connection.rb:34:in `block in with_lock'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activesupport-7.0.8.1/lib/active_support/concurrency/load_interlock_aware_monitor.rb:25:in `handle_interrupt'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activesupport-7.0.8.1/lib/active_support/concurrency/load_interlock_aware_monitor.rb:25:in `block in synchronize'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activesupport-7.0.8.1/lib/active_support/concurrency/load_interlock_aware_monitor.rb:21:in `handle_interrupt'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/activesupport-7.0.8.1/lib/active_support/concurrency/load_interlock_aware_monitor.rb:21:in `synchronize'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/mini_sql-1.5.0/lib/mini_sql/active_record_postgres/connection.rb:34:in `with_lock'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/mini_sql-1.5.0/lib/mini_sql/active_record_postgres/connection.rb:38:in `run'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/mini_sql-1.5.0/lib/mini_sql/postgres/connection.rb:99:in `query'
/var/www/discourse/plugins/discourse-ai/lib/embeddings/vector_representations/base.rb:272:in `asymmetric_rag_fragment_similarity_search'
/var/www/discourse/plugins/discourse-ai/lib/ai_bot/personas/persona.rb:286:in `rag_fragments_prompt'
/var/www/discourse/plugins/discourse-ai/lib/ai_bot/personas/persona.rb:156:in `craft_prompt'
/var/www/discourse/plugins/discourse-ai/lib/ai_bot/bot.rb:54:in `reply'
/var/www/discourse/plugins/discourse-ai/lib/ai_bot/playground.rb:424: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:305:in `block (2 levels) in perform'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/rails_multisite-6.0.0/lib/rails_multisite/connection_management/null_instance.rb:49:in `with_connection'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/rails_multisite-6.0.0/lib/rails_multisite/connection_management.rb:21:in `with_connection'
/var/www/discourse/app/jobs/base.rb:292:in `block in perform'
/var/www/discourse/app/jobs/base.rb:288:in `each'
/var/www/discourse/app/jobs/base.rb:288:in `perform'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:202:in `execute_job'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:170:in `block (2 levels) in process'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/middleware/chain.rb:177:in `block in invoke'
/var/www/discourse/lib/sidekiq/pausable.rb:132:in `call'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/middleware/chain.rb:179:in `block in invoke'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/middleware/chain.rb:182:in `invoke'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:169:in `block in process'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:136:in `block (6 levels) in dispatch'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/job_retry.rb:113:in `local'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:135:in `block (5 levels) in dispatch'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq.rb:44:in `block in <module:Sidekiq>'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:131:in `block (4 levels) in dispatch'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:263:in `stats'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:126:in `block (3 levels) in dispatch'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/job_logger.rb:13:in `call'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:125:in `block (2 levels) in dispatch'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/job_retry.rb:80:in `global'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:124:in `block in dispatch'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/job_logger.rb:39:in `prepare'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:123:in `dispatch'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:168:in `process'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:78:in `process_one'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/processor.rb:68:in `run'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/component.rb:8:in `watchdog'
/var/www/discourse/vendor/bundle/ruby/3.2.0/gems/sidekiq-6.5.12/lib/sidekiq/component.rb:17:in `block in safe_thread'

[sam (Sam Saffron)]

Obrigado por levantar a questão, vamos dar uma olhada!

[Falco]

Qual é a saída da execução dos seguintes comandos em um console Rails?

strategy = DiscourseAi::Embeddings::Strategies::Truncation.new
vector_rep = DiscourseAi::Embeddings::VectorRepresentations::Base.current_representation(strategy)
vector_rep.vector_from("test")

Além disso, nossa API foi projetada para funcionar contra alguém executando o GitHub - huggingface/text-embeddings-inference: A blazing fast inference solution for text embeddings models por conta própria, conforme a documentação, então é possível que não funcione contra a versão hospedada.

Se você fornecer o backtrace, poderemos investigar para fazê-lo funcionar.

[StevePlex (Steve King)]

@Falco

Aqui está o que aconteceu quando executei o código de teste (com bge-large-en rodando em uma instância de endpoint dedicado da AWS configurada como modelo de incorporação)

---

root@studyqa-ubuntu-s-1vcpu-2gb-amd-sfo3-01-app:/var/www/discourse# rails c

[1] pry(main)> strategy = DiscourseAi::Embeddings::Strategies::Truncation.new

puts "Strategy initialized"

vector_rep = DiscourseAi::Embeddings::VectorRepresentations::Base.current_representation(strategy)

puts "Vector representation obtained"

vector = vector_rep.vector_from("test")

[1] pry(main)> strategy = DiscourseAi::Embeddings::Strategies::Truncation.new

puts "Strategy initialized"

vector_rep = DiscourseAi::Embeddings::VectorRepresentations::Base.current_representation(strategy)

puts "Vector representation obtained"

vector = vector_rep.vector_from("test")

puts "Vector generated"

puts vector.inspect

Strategy initialized

Vector representation obtained

Vector generated

[:embeddings, [-0.0020444912370294333, 0.008787356317043304, -0.010865539312362671, 0.01865551434457302, -0.02099628746509552, -0.009864491410553455, -0.0011329081607982516, 0.02949545904994011, 0.027839021757245064, 0.043966952711343765, 0.0406080037355423, 0.0016647017328068614, 0.007204003632068634, -0.03770752251148224, -0.025242917239665985, -0.0015279072104021907, -0.02805529721081257, -0.020901955664157867, -0.029206447303295135, -0.006209365092217922, -0.02105099707841873,

etc.

---

Parece que está atingindo o bge-large-en na AWS:

- 2024-05-29T13:57:34.609+00:00 Batches: 0%| | 0/1 [00:00<?, ?it/s] Batches: 100%|██████████| 1/1 [00:00<00:00, 4.80it/s] Batches: 100%|██████████| 1/1 [00:00<00:00, 4.79it/s]

• 2024/05/29 09:57:34
INFO | POST / | Duration: 212.84 ms


- 2024-05-29T13:57:53.806+00:00 Batches: 0%| | 0/1 [00:00<?, ?it/s] Batches: 100%|██████████| 1/1 [00:01<00:00, 1.97s/it] Batches: 100%|██████████| 1/1 [00:01<00:00, 1.97s/it]

• 2024/05/29 09:57:53
INFO | POST / | Duration: 1978.36 ms

[Falco]

Parece que está funcionando bem?

Talvez o problema seja o re-ranker? Você pode desativar o ai_hugging_face_tei_reranker_endpoint e testar se o RAG funciona?

[StevePlex (Steve King)]

reranker desligado.. ainda sem embedding.. recebendo esta mensagem em ambas as extremidades:

---

Teste de execução do Discourse LLM:

Tentar contatar o modelo retornou este erro: {“error”:“Body needs to provide a inputs key, recieved: b’{\"model\":\"bge-large-en\",\"temperature\":0.7,\"messages\":[{\"role\":\"system\",\"content\":\"You are a helpful bot\"},{\"role\":\"user\",\"content\":\"How much is 1 + 1?\"}],\"max_tokens\":1009}'”}

---

Log do bge-large-en

• 2024/05/29 13:40:03

ERROR | Body needs to provide a inputs key, recieved: b’{"model":"bge-large-en","temperature":0.7,"messages":[{"role":"system","content":"You are a helpful bot"},{"role":"user","content":"How much is 1 + 1?"}],"max_tokens":1009}’

---

discourse b1b218aa99
discourse-ai d812ecf5

[sam (Sam Saffron)]

@StevePlex

Ao tentar contatar o modelo, recebi este erro: {“error”:“Body needs to provide a inputs key, recieved: b’{"model":"bge-large-en","temperature":0.7,"messages":[{"role":"system","content":"You are a helpful bot"},{"role":"user","content":"How much is 1 + 1?"}],"max_tokens":1009}'”}

Não é assim que devemos testar embeddings esse é um teste de LLM, não um teste de modelo de embedding que esperaria números de volta. A interface do usuário do LLM não é onde você adicionaria isso; precisaríamos de uma interface de usuário de Embedding para isso, que ainda não temos. Modelos de embedding são configurados apenas nas configurações do site.

[StevePlex (Steve King)]

Sim. Faz sentido.

(Tentei observar que eu estava usando apenas o teste de execução do LLM para confirmar a “conectividade” (veja abaixo)! Deveria ter deixado isso mais claro.)

---

---