discourse-ai/spec/lib/modules/ai_bot/personas/persona_spec.rb

479 lines
15 KiB
Ruby
Raw Normal View History

#frozen_string_literal: true
class TestPersona < DiscourseAi::AiBot::Personas::Persona
def tools
[
DiscourseAi::AiBot::Tools::ListTags,
DiscourseAi::AiBot::Tools::Search,
DiscourseAi::AiBot::Tools::Image,
]
end
def system_prompt
<<~PROMPT
{site_url}
{site_title}
{site_description}
{participants}
{time}
PROMPT
end
end
RSpec.describe DiscourseAi::AiBot::Personas::Persona do
let :persona do
TestPersona.new
end
let :topic_with_users do
topic = Topic.new
topic.allowed_users = [User.new(username: "joe"), User.new(username: "jane")]
topic
end
after do
# we are rolling back transactions so we can create poison cache
AiPersona.persona_cache.flush!
end
let(:context) do
{
site_url: Discourse.base_url,
site_title: "test site title",
site_description: "test site description",
time: Time.zone.now,
participants: topic_with_users.allowed_users.map(&:username).join(", "),
}
end
FEATURE: AI artifacts (#898) 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.
2024-11-18 17:22:39 -05:00
fab!(:admin)
fab!(:user)
FEATURE: Add Question Consolidator for robust Upload support in Personas (#596) This commit introduces a new feature for AI Personas called the "Question Consolidator LLM". The purpose of the Question Consolidator is to consolidate a user's latest question into a self-contained, context-rich question before querying the vector database for relevant fragments. This helps improve the quality and relevance of the retrieved fragments. Previous to this change we used the last 10 interactions, this is not ideal cause the RAG would "lock on" to an answer. EG: - User: how many cars are there in europe - Model: detailed answer about cars in europe including the term car and vehicle many times - User: Nice, what about trains are there in the US In the above example "trains" and "US" becomes very low signal given there are pages and pages talking about cars and europe. This mean retrieval is sub optimal. Instead, we pass the history to the "question consolidator", it would simply consolidate the question to "How many trains are there in the United States", which would make it fare easier for the vector db to find relevant content. The llm used for question consolidator can often be less powerful than the model you are talking to, we recommend using lighter weight and fast models cause the task is very simple. This is configurable from the persona ui. This PR also removes support for {uploads} placeholder, this is too complicated to get right and we want freedom to shift RAG implementation. Key changes: 1. Added a new `question_consolidator_llm` column to the `ai_personas` table to store the LLM model used for question consolidation. 2. Implemented the `QuestionConsolidator` module which handles the logic for consolidating the user's latest question. It extracts the relevant user and model messages from the conversation history, truncates them if needed to fit within the token limit, and generates a consolidated question prompt. 3. Updated the `Persona` class to use the Question Consolidator LLM (if configured) when crafting the RAG fragments prompt. It passes the conversation context to the consolidator to generate a self-contained question. 4. Added UI elements in the AI Persona editor to allow selecting the Question Consolidator LLM. Also made some UI tweaks to conditionally show/hide certain options based on persona configuration. 5. Wrote unit tests for the QuestionConsolidator module and updated existing persona tests to cover the new functionality. This feature enables AI Personas to better understand the context and intent behind a user's question by consolidating the conversation history into a single, focused question. This can lead to more relevant and accurate responses from the AI assistant.
2024-04-29 23:49:21 -04:00
fab!(:upload)
it "renders the system prompt" do
freeze_time
rendered = persona.craft_prompt(context)
system_message = rendered.messages.first[:content]
expect(system_message).to include(Discourse.base_url)
expect(system_message).to include("test site title")
expect(system_message).to include("test site description")
expect(system_message).to include("joe, jane")
expect(system_message).to include(Time.zone.now.to_s)
tools = rendered.tools
expect(tools.find { |t| t[:name] == "search" }).to be_present
expect(tools.find { |t| t[:name] == "tags" }).to be_present
# needs to be configured so it is not available
expect(tools.find { |t| t[:name] == "image" }).to be_nil
end
it "can parse string that are wrapped in quotes" do
SiteSetting.ai_stability_api_key = "123"
tool_call =
DiscourseAi::Completions::ToolCall.new(
name: "image",
id: "call_JtYQMful5QKqw97XFsHzPweB",
parameters: {
prompts: ["cat oil painting", "big car"],
aspect_ratio: "16:9",
},
)
tool_instance =
DiscourseAi::AiBot::Personas::Artist.new.find_tool(
tool_call,
bot_user: nil,
llm: nil,
context: nil,
)
expect(tool_instance.parameters[:prompts]).to eq(["cat oil painting", "big car"])
expect(tool_instance.parameters[:aspect_ratio]).to eq("16:9")
end
it "enforces enums" do
tool_call =
DiscourseAi::Completions::ToolCall.new(
name: "search",
id: "call_JtYQMful5QKqw97XFsHzPweB",
parameters: {
max_posts: "3.2",
status: "cow",
foo: "bar",
},
)
tool_instance =
DiscourseAi::AiBot::Personas::General.new.find_tool(
tool_call,
bot_user: nil,
llm: nil,
context: nil,
)
expect(tool_instance.parameters.key?(:status)).to eq(false)
tool_call =
DiscourseAi::Completions::ToolCall.new(
name: "search",
id: "call_JtYQMful5QKqw97XFsHzPweB",
parameters: {
max_posts: "3.2",
status: "open",
foo: "bar",
},
)
tool_instance =
DiscourseAi::AiBot::Personas::General.new.find_tool(
tool_call,
bot_user: nil,
llm: nil,
context: nil,
)
expect(tool_instance.parameters[:status]).to eq("open")
end
it "can coerce integers" do
tool_call =
DiscourseAi::Completions::ToolCall.new(
name: "search",
id: "call_JtYQMful5QKqw97XFsHzPweB",
parameters: {
max_posts: "3.2",
search_query: "hello world",
foo: "bar",
},
)
search =
DiscourseAi::AiBot::Personas::General.new.find_tool(
tool_call,
bot_user: nil,
llm: nil,
context: nil,
)
expect(search.parameters[:max_posts]).to eq(3)
expect(search.parameters[:search_query]).to eq("hello world")
expect(search.parameters.key?(:foo)).to eq(false)
end
it "can correctly parse arrays in tools" do
SiteSetting.ai_openai_api_key = "123"
tool_call =
DiscourseAi::Completions::ToolCall.new(
name: "dall_e",
id: "call_JtYQMful5QKqw97XFsHzPweB",
parameters: {
prompts: ["cat oil painting", "big car"],
},
)
tool_instance =
DiscourseAi::AiBot::Personas::DallE3.new.find_tool(
tool_call,
bot_user: nil,
llm: nil,
context: nil,
)
expect(tool_instance.parameters[:prompts]).to eq(["cat oil painting", "big car"])
end
describe "custom personas" do
it "is able to find custom personas" do
Group.refresh_automatic_groups!
# define an ai persona everyone can see
persona =
AiPersona.create!(
name: "zzzpun_bot",
description: "you write puns",
system_prompt: "you are pun bot",
tools: ["Image"],
allowed_group_ids: [Group::AUTO_GROUPS[:trust_level_0]],
)
FEATURE: UI to update ai personas on admin page (#290) Introduces a UI to manage customizable personas (admin only feature) Part of the change was some extensive internal refactoring: - AIBot now has a persona set in the constructor, once set it never changes - Command now takes in bot as a constructor param, so it has the correct persona and is not generating AIBot objects on the fly - Added a .prettierignore file, due to the way ALE is configured in nvim it is a pre-req for prettier to work - Adds a bunch of validations on the AIPersona model, system personas (artist/creative etc...) are all seeded. We now ensure - name uniqueness, and only allow certain properties to be touched for system personas. - (JS note) the client side design takes advantage of nested routes, the parent route for personas gets all the personas via this.store.findAll("ai-persona") then child routes simply reach into this model to find a particular persona. - (JS note) data is sideloaded into the ai-persona model the meta property supplied from the controller, resultSetMeta - This removes ai_bot_enabled_personas and ai_bot_enabled_chat_commands, both should be controlled from the UI on a per persona basis - Fixes a long standing bug in token accounting ... we were doing to_json.length instead of to_json.to_s.length - Amended it so {commands} are always inserted at the end unconditionally, no need to add it to the template of the system message as it just confuses things - Adds a concept of required_commands to stock personas, these are commands that must be configured for this stock persona to show up. - Refactored tests so we stop requiring inference_stubs, it was very confusing to need it, added to plugin.rb for now which at least is clearer - Migrates the persona selector to gjs --------- Co-authored-by: Joffrey JAFFEUX <j.jaffeux@gmail.com> Co-authored-by: Martin Brennan <martin@discourse.org>
2023-11-21 00:56:43 -05:00
custom_persona = DiscourseAi::AiBot::Personas::Persona.all(user: user).last
expect(custom_persona.name).to eq("zzzpun_bot")
expect(custom_persona.description).to eq("you write puns")
FEATURE: UI to update ai personas on admin page (#290) Introduces a UI to manage customizable personas (admin only feature) Part of the change was some extensive internal refactoring: - AIBot now has a persona set in the constructor, once set it never changes - Command now takes in bot as a constructor param, so it has the correct persona and is not generating AIBot objects on the fly - Added a .prettierignore file, due to the way ALE is configured in nvim it is a pre-req for prettier to work - Adds a bunch of validations on the AIPersona model, system personas (artist/creative etc...) are all seeded. We now ensure - name uniqueness, and only allow certain properties to be touched for system personas. - (JS note) the client side design takes advantage of nested routes, the parent route for personas gets all the personas via this.store.findAll("ai-persona") then child routes simply reach into this model to find a particular persona. - (JS note) data is sideloaded into the ai-persona model the meta property supplied from the controller, resultSetMeta - This removes ai_bot_enabled_personas and ai_bot_enabled_chat_commands, both should be controlled from the UI on a per persona basis - Fixes a long standing bug in token accounting ... we were doing to_json.length instead of to_json.to_s.length - Amended it so {commands} are always inserted at the end unconditionally, no need to add it to the template of the system message as it just confuses things - Adds a concept of required_commands to stock personas, these are commands that must be configured for this stock persona to show up. - Refactored tests so we stop requiring inference_stubs, it was very confusing to need it, added to plugin.rb for now which at least is clearer - Migrates the persona selector to gjs --------- Co-authored-by: Joffrey JAFFEUX <j.jaffeux@gmail.com> Co-authored-by: Martin Brennan <martin@discourse.org>
2023-11-21 00:56:43 -05:00
instance = custom_persona.new
expect(instance.tools).to eq([DiscourseAi::AiBot::Tools::Image])
expect(instance.craft_prompt(context).messages.first[:content]).to eq("you are pun bot")
FEATURE: UI to update ai personas on admin page (#290) Introduces a UI to manage customizable personas (admin only feature) Part of the change was some extensive internal refactoring: - AIBot now has a persona set in the constructor, once set it never changes - Command now takes in bot as a constructor param, so it has the correct persona and is not generating AIBot objects on the fly - Added a .prettierignore file, due to the way ALE is configured in nvim it is a pre-req for prettier to work - Adds a bunch of validations on the AIPersona model, system personas (artist/creative etc...) are all seeded. We now ensure - name uniqueness, and only allow certain properties to be touched for system personas. - (JS note) the client side design takes advantage of nested routes, the parent route for personas gets all the personas via this.store.findAll("ai-persona") then child routes simply reach into this model to find a particular persona. - (JS note) data is sideloaded into the ai-persona model the meta property supplied from the controller, resultSetMeta - This removes ai_bot_enabled_personas and ai_bot_enabled_chat_commands, both should be controlled from the UI on a per persona basis - Fixes a long standing bug in token accounting ... we were doing to_json.length instead of to_json.to_s.length - Amended it so {commands} are always inserted at the end unconditionally, no need to add it to the template of the system message as it just confuses things - Adds a concept of required_commands to stock personas, these are commands that must be configured for this stock persona to show up. - Refactored tests so we stop requiring inference_stubs, it was very confusing to need it, added to plugin.rb for now which at least is clearer - Migrates the persona selector to gjs --------- Co-authored-by: Joffrey JAFFEUX <j.jaffeux@gmail.com> Co-authored-by: Martin Brennan <martin@discourse.org>
2023-11-21 00:56:43 -05:00
# should update
persona.update!(name: "zzzpun_bot2")
custom_persona = DiscourseAi::AiBot::Personas::Persona.all(user: user).last
expect(custom_persona.name).to eq("zzzpun_bot2")
# can be disabled
persona.update!(enabled: false)
last_persona = DiscourseAi::AiBot::Personas::Persona.all(user: user).last
expect(last_persona.name).not_to eq("zzzpun_bot2")
persona.update!(enabled: true)
# no groups have access
persona.update!(allowed_group_ids: [])
FEATURE: UI to update ai personas on admin page (#290) Introduces a UI to manage customizable personas (admin only feature) Part of the change was some extensive internal refactoring: - AIBot now has a persona set in the constructor, once set it never changes - Command now takes in bot as a constructor param, so it has the correct persona and is not generating AIBot objects on the fly - Added a .prettierignore file, due to the way ALE is configured in nvim it is a pre-req for prettier to work - Adds a bunch of validations on the AIPersona model, system personas (artist/creative etc...) are all seeded. We now ensure - name uniqueness, and only allow certain properties to be touched for system personas. - (JS note) the client side design takes advantage of nested routes, the parent route for personas gets all the personas via this.store.findAll("ai-persona") then child routes simply reach into this model to find a particular persona. - (JS note) data is sideloaded into the ai-persona model the meta property supplied from the controller, resultSetMeta - This removes ai_bot_enabled_personas and ai_bot_enabled_chat_commands, both should be controlled from the UI on a per persona basis - Fixes a long standing bug in token accounting ... we were doing to_json.length instead of to_json.to_s.length - Amended it so {commands} are always inserted at the end unconditionally, no need to add it to the template of the system message as it just confuses things - Adds a concept of required_commands to stock personas, these are commands that must be configured for this stock persona to show up. - Refactored tests so we stop requiring inference_stubs, it was very confusing to need it, added to plugin.rb for now which at least is clearer - Migrates the persona selector to gjs --------- Co-authored-by: Joffrey JAFFEUX <j.jaffeux@gmail.com> Co-authored-by: Martin Brennan <martin@discourse.org>
2023-11-21 00:56:43 -05:00
last_persona = DiscourseAi::AiBot::Personas::Persona.all(user: user).last
expect(last_persona.name).not_to eq("zzzpun_bot2")
end
end
describe "available personas" do
it "includes all personas by default" do
Group.refresh_automatic_groups!
# must be enabled to see it
SiteSetting.ai_stability_api_key = "abc"
SiteSetting.ai_google_custom_search_api_key = "abc"
SiteSetting.ai_google_custom_search_cx = "abc123"
# should be ordered by priority and then alpha
expect(DiscourseAi::AiBot::Personas::Persona.all(user: user)).to eq(
[
DiscourseAi::AiBot::Personas::General,
DiscourseAi::AiBot::Personas::Artist,
DiscourseAi::AiBot::Personas::Creative,
DiscourseAi::AiBot::Personas::DiscourseHelper,
DiscourseAi::AiBot::Personas::GithubHelper,
DiscourseAi::AiBot::Personas::Researcher,
DiscourseAi::AiBot::Personas::SettingsExplorer,
DiscourseAi::AiBot::Personas::SqlHelper,
],
)
FEATURE: AI artifacts (#898) 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.
2024-11-18 17:22:39 -05:00
# it should allow staff access to WebArtifactCreator
expect(DiscourseAi::AiBot::Personas::Persona.all(user: admin)).to eq(
[
DiscourseAi::AiBot::Personas::General,
DiscourseAi::AiBot::Personas::Artist,
DiscourseAi::AiBot::Personas::Creative,
DiscourseAi::AiBot::Personas::DiscourseHelper,
DiscourseAi::AiBot::Personas::GithubHelper,
DiscourseAi::AiBot::Personas::Researcher,
DiscourseAi::AiBot::Personas::SettingsExplorer,
DiscourseAi::AiBot::Personas::SqlHelper,
DiscourseAi::AiBot::Personas::WebArtifactCreator,
],
)
# omits personas if key is missing
SiteSetting.ai_stability_api_key = ""
SiteSetting.ai_google_custom_search_api_key = ""
FEATURE: AI artifacts (#898) 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.
2024-11-18 17:22:39 -05:00
SiteSetting.ai_artifact_security = "disabled"
FEATURE: AI artifacts (#898) 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.
2024-11-18 17:22:39 -05:00
expect(DiscourseAi::AiBot::Personas::Persona.all(user: admin)).to contain_exactly(
DiscourseAi::AiBot::Personas::General,
DiscourseAi::AiBot::Personas::SqlHelper,
DiscourseAi::AiBot::Personas::SettingsExplorer,
DiscourseAi::AiBot::Personas::Creative,
DiscourseAi::AiBot::Personas::DiscourseHelper,
DiscourseAi::AiBot::Personas::GithubHelper,
)
AiPersona.find(
DiscourseAi::AiBot::Personas::Persona.system_personas[
DiscourseAi::AiBot::Personas::General
],
).update!(enabled: false)
expect(DiscourseAi::AiBot::Personas::Persona.all(user: user)).to contain_exactly(
DiscourseAi::AiBot::Personas::SqlHelper,
DiscourseAi::AiBot::Personas::SettingsExplorer,
DiscourseAi::AiBot::Personas::Creative,
DiscourseAi::AiBot::Personas::DiscourseHelper,
DiscourseAi::AiBot::Personas::GithubHelper,
)
end
end
describe "#craft_prompt" do
before do
Group.refresh_automatic_groups!
SiteSetting.ai_embeddings_discourse_service_api_endpoint = "http://test.com"
SiteSetting.ai_embeddings_enabled = true
end
let(:ai_persona) { DiscourseAi::AiBot::Personas::Persona.all(user: user).first.new }
let(:with_cc) do
context.merge(conversation_context: [{ content: "Tell me the time", type: :user }])
end
context "when a persona has no uploads" do
it "doesn't include RAG guidance" do
guidance_fragment =
"The following texts will give you additional guidance to elaborate a response."
expect(ai_persona.craft_prompt(with_cc).messages.first[:content]).not_to include(
guidance_fragment,
)
end
end
FEATURE: Add Question Consolidator for robust Upload support in Personas (#596) This commit introduces a new feature for AI Personas called the "Question Consolidator LLM". The purpose of the Question Consolidator is to consolidate a user's latest question into a self-contained, context-rich question before querying the vector database for relevant fragments. This helps improve the quality and relevance of the retrieved fragments. Previous to this change we used the last 10 interactions, this is not ideal cause the RAG would "lock on" to an answer. EG: - User: how many cars are there in europe - Model: detailed answer about cars in europe including the term car and vehicle many times - User: Nice, what about trains are there in the US In the above example "trains" and "US" becomes very low signal given there are pages and pages talking about cars and europe. This mean retrieval is sub optimal. Instead, we pass the history to the "question consolidator", it would simply consolidate the question to "How many trains are there in the United States", which would make it fare easier for the vector db to find relevant content. The llm used for question consolidator can often be less powerful than the model you are talking to, we recommend using lighter weight and fast models cause the task is very simple. This is configurable from the persona ui. This PR also removes support for {uploads} placeholder, this is too complicated to get right and we want freedom to shift RAG implementation. Key changes: 1. Added a new `question_consolidator_llm` column to the `ai_personas` table to store the LLM model used for question consolidation. 2. Implemented the `QuestionConsolidator` module which handles the logic for consolidating the user's latest question. It extracts the relevant user and model messages from the conversation history, truncates them if needed to fit within the token limit, and generates a consolidated question prompt. 3. Updated the `Persona` class to use the Question Consolidator LLM (if configured) when crafting the RAG fragments prompt. It passes the conversation context to the consolidator to generate a self-contained question. 4. Added UI elements in the AI Persona editor to allow selecting the Question Consolidator LLM. Also made some UI tweaks to conditionally show/hide certain options based on persona configuration. 5. Wrote unit tests for the QuestionConsolidator module and updated existing persona tests to cover the new functionality. This feature enables AI Personas to better understand the context and intent behind a user's question by consolidating the conversation history into a single, focused question. This can lead to more relevant and accurate responses from the AI assistant.
2024-04-29 23:49:21 -04:00
context "when RAG is running with a question consolidator" do
let(:consolidated_question) { "what is the time in france?" }
fab!(:llm_model) { Fabricate(:fake_model) }
FEATURE: Add Question Consolidator for robust Upload support in Personas (#596) This commit introduces a new feature for AI Personas called the "Question Consolidator LLM". The purpose of the Question Consolidator is to consolidate a user's latest question into a self-contained, context-rich question before querying the vector database for relevant fragments. This helps improve the quality and relevance of the retrieved fragments. Previous to this change we used the last 10 interactions, this is not ideal cause the RAG would "lock on" to an answer. EG: - User: how many cars are there in europe - Model: detailed answer about cars in europe including the term car and vehicle many times - User: Nice, what about trains are there in the US In the above example "trains" and "US" becomes very low signal given there are pages and pages talking about cars and europe. This mean retrieval is sub optimal. Instead, we pass the history to the "question consolidator", it would simply consolidate the question to "How many trains are there in the United States", which would make it fare easier for the vector db to find relevant content. The llm used for question consolidator can often be less powerful than the model you are talking to, we recommend using lighter weight and fast models cause the task is very simple. This is configurable from the persona ui. This PR also removes support for {uploads} placeholder, this is too complicated to get right and we want freedom to shift RAG implementation. Key changes: 1. Added a new `question_consolidator_llm` column to the `ai_personas` table to store the LLM model used for question consolidation. 2. Implemented the `QuestionConsolidator` module which handles the logic for consolidating the user's latest question. It extracts the relevant user and model messages from the conversation history, truncates them if needed to fit within the token limit, and generates a consolidated question prompt. 3. Updated the `Persona` class to use the Question Consolidator LLM (if configured) when crafting the RAG fragments prompt. It passes the conversation context to the consolidator to generate a self-contained question. 4. Added UI elements in the AI Persona editor to allow selecting the Question Consolidator LLM. Also made some UI tweaks to conditionally show/hide certain options based on persona configuration. 5. Wrote unit tests for the QuestionConsolidator module and updated existing persona tests to cover the new functionality. This feature enables AI Personas to better understand the context and intent behind a user's question by consolidating the conversation history into a single, focused question. This can lead to more relevant and accurate responses from the AI assistant.
2024-04-29 23:49:21 -04:00
it "will run the question consolidator" do
strategy = DiscourseAi::Embeddings::Strategies::Truncation.new
vector_rep =
DiscourseAi::Embeddings::VectorRepresentations::Base.current_representation(strategy)
context_embedding = vector_rep.dimensions.times.map { rand(-1.0...1.0) }
FEATURE: Add Question Consolidator for robust Upload support in Personas (#596) This commit introduces a new feature for AI Personas called the "Question Consolidator LLM". The purpose of the Question Consolidator is to consolidate a user's latest question into a self-contained, context-rich question before querying the vector database for relevant fragments. This helps improve the quality and relevance of the retrieved fragments. Previous to this change we used the last 10 interactions, this is not ideal cause the RAG would "lock on" to an answer. EG: - User: how many cars are there in europe - Model: detailed answer about cars in europe including the term car and vehicle many times - User: Nice, what about trains are there in the US In the above example "trains" and "US" becomes very low signal given there are pages and pages talking about cars and europe. This mean retrieval is sub optimal. Instead, we pass the history to the "question consolidator", it would simply consolidate the question to "How many trains are there in the United States", which would make it fare easier for the vector db to find relevant content. The llm used for question consolidator can often be less powerful than the model you are talking to, we recommend using lighter weight and fast models cause the task is very simple. This is configurable from the persona ui. This PR also removes support for {uploads} placeholder, this is too complicated to get right and we want freedom to shift RAG implementation. Key changes: 1. Added a new `question_consolidator_llm` column to the `ai_personas` table to store the LLM model used for question consolidation. 2. Implemented the `QuestionConsolidator` module which handles the logic for consolidating the user's latest question. It extracts the relevant user and model messages from the conversation history, truncates them if needed to fit within the token limit, and generates a consolidated question prompt. 3. Updated the `Persona` class to use the Question Consolidator LLM (if configured) when crafting the RAG fragments prompt. It passes the conversation context to the consolidator to generate a self-contained question. 4. Added UI elements in the AI Persona editor to allow selecting the Question Consolidator LLM. Also made some UI tweaks to conditionally show/hide certain options based on persona configuration. 5. Wrote unit tests for the QuestionConsolidator module and updated existing persona tests to cover the new functionality. This feature enables AI Personas to better understand the context and intent behind a user's question by consolidating the conversation history into a single, focused question. This can lead to more relevant and accurate responses from the AI assistant.
2024-04-29 23:49:21 -04:00
EmbeddingsGenerationStubs.discourse_service(
SiteSetting.ai_embeddings_model,
consolidated_question,
context_embedding,
)
FEATURE: Add Question Consolidator for robust Upload support in Personas (#596) This commit introduces a new feature for AI Personas called the "Question Consolidator LLM". The purpose of the Question Consolidator is to consolidate a user's latest question into a self-contained, context-rich question before querying the vector database for relevant fragments. This helps improve the quality and relevance of the retrieved fragments. Previous to this change we used the last 10 interactions, this is not ideal cause the RAG would "lock on" to an answer. EG: - User: how many cars are there in europe - Model: detailed answer about cars in europe including the term car and vehicle many times - User: Nice, what about trains are there in the US In the above example "trains" and "US" becomes very low signal given there are pages and pages talking about cars and europe. This mean retrieval is sub optimal. Instead, we pass the history to the "question consolidator", it would simply consolidate the question to "How many trains are there in the United States", which would make it fare easier for the vector db to find relevant content. The llm used for question consolidator can often be less powerful than the model you are talking to, we recommend using lighter weight and fast models cause the task is very simple. This is configurable from the persona ui. This PR also removes support for {uploads} placeholder, this is too complicated to get right and we want freedom to shift RAG implementation. Key changes: 1. Added a new `question_consolidator_llm` column to the `ai_personas` table to store the LLM model used for question consolidation. 2. Implemented the `QuestionConsolidator` module which handles the logic for consolidating the user's latest question. It extracts the relevant user and model messages from the conversation history, truncates them if needed to fit within the token limit, and generates a consolidated question prompt. 3. Updated the `Persona` class to use the Question Consolidator LLM (if configured) when crafting the RAG fragments prompt. It passes the conversation context to the consolidator to generate a self-contained question. 4. Added UI elements in the AI Persona editor to allow selecting the Question Consolidator LLM. Also made some UI tweaks to conditionally show/hide certain options based on persona configuration. 5. Wrote unit tests for the QuestionConsolidator module and updated existing persona tests to cover the new functionality. This feature enables AI Personas to better understand the context and intent behind a user's question by consolidating the conversation history into a single, focused question. This can lead to more relevant and accurate responses from the AI assistant.
2024-04-29 23:49:21 -04:00
custom_ai_persona =
Fabricate(
:ai_persona,
name: "custom",
rag_conversation_chunks: 3,
allowed_group_ids: [Group::AUTO_GROUPS[:trust_level_0]],
question_consolidator_llm: "custom:#{llm_model.id}",
FEATURE: Add Question Consolidator for robust Upload support in Personas (#596) This commit introduces a new feature for AI Personas called the "Question Consolidator LLM". The purpose of the Question Consolidator is to consolidate a user's latest question into a self-contained, context-rich question before querying the vector database for relevant fragments. This helps improve the quality and relevance of the retrieved fragments. Previous to this change we used the last 10 interactions, this is not ideal cause the RAG would "lock on" to an answer. EG: - User: how many cars are there in europe - Model: detailed answer about cars in europe including the term car and vehicle many times - User: Nice, what about trains are there in the US In the above example "trains" and "US" becomes very low signal given there are pages and pages talking about cars and europe. This mean retrieval is sub optimal. Instead, we pass the history to the "question consolidator", it would simply consolidate the question to "How many trains are there in the United States", which would make it fare easier for the vector db to find relevant content. The llm used for question consolidator can often be less powerful than the model you are talking to, we recommend using lighter weight and fast models cause the task is very simple. This is configurable from the persona ui. This PR also removes support for {uploads} placeholder, this is too complicated to get right and we want freedom to shift RAG implementation. Key changes: 1. Added a new `question_consolidator_llm` column to the `ai_personas` table to store the LLM model used for question consolidation. 2. Implemented the `QuestionConsolidator` module which handles the logic for consolidating the user's latest question. It extracts the relevant user and model messages from the conversation history, truncates them if needed to fit within the token limit, and generates a consolidated question prompt. 3. Updated the `Persona` class to use the Question Consolidator LLM (if configured) when crafting the RAG fragments prompt. It passes the conversation context to the consolidator to generate a self-contained question. 4. Added UI elements in the AI Persona editor to allow selecting the Question Consolidator LLM. Also made some UI tweaks to conditionally show/hide certain options based on persona configuration. 5. Wrote unit tests for the QuestionConsolidator module and updated existing persona tests to cover the new functionality. This feature enables AI Personas to better understand the context and intent behind a user's question by consolidating the conversation history into a single, focused question. This can lead to more relevant and accurate responses from the AI assistant.
2024-04-29 23:49:21 -04:00
)
UploadReference.ensure_exist!(target: custom_ai_persona, upload_ids: [upload.id])
custom_persona =
DiscourseAi::AiBot::Personas::Persona.find_by(id: custom_ai_persona.id, user: user).new
# this means that we will consolidate
ctx =
with_cc.merge(
conversation_context: [
{ content: "Tell me the time", type: :user },
{ content: "the time is 1", type: :model },
{ content: "in france?", type: :user },
],
)
DiscourseAi::Completions::Endpoints::Fake.with_fake_content(consolidated_question) do
custom_persona.craft_prompt(ctx).messages.first[:content]
end
message =
DiscourseAi::Completions::Endpoints::Fake.last_call[:dialect].prompt.messages.last[
:content
]
expect(message).to include("Tell me the time")
expect(message).to include("the time is 1")
expect(message).to include("in france?")
end
end
context "when a persona has RAG uploads" do
def stub_fragments(limit, expected_limit: nil)
candidate_ids = []
limit.times do |i|
candidate_ids << Fabricate(
:rag_document_fragment,
fragment: "fragment-n#{i}",
target_id: ai_persona.id,
target_type: "AiPersona",
upload: upload,
).id
end
DiscourseAi::Embeddings::VectorRepresentations::BgeLargeEn
.any_instance
.expects(:asymmetric_rag_fragment_similarity_search)
.with { |args, kwargs| kwargs[:limit] == (expected_limit || limit) }
.returns(candidate_ids)
end
before do
stored_ai_persona = AiPersona.find(ai_persona.id)
UploadReference.ensure_exist!(target: stored_ai_persona, upload_ids: [upload.id])
context_embedding = [0.049382, 0.9999]
EmbeddingsGenerationStubs.discourse_service(
SiteSetting.ai_embeddings_model,
with_cc.dig(:conversation_context, 0, :content),
context_embedding,
)
end
context "when persona allows for less fragments" do
before { stub_fragments(3) }
it "will only pick 3 fragments" do
custom_ai_persona =
Fabricate(
:ai_persona,
name: "custom",
rag_conversation_chunks: 3,
allowed_group_ids: [Group::AUTO_GROUPS[:trust_level_0]],
)
UploadReference.ensure_exist!(target: custom_ai_persona, upload_ids: [upload.id])
custom_persona =
DiscourseAi::AiBot::Personas::Persona.find_by(id: custom_ai_persona.id, user: user).new
expect(custom_persona.class.rag_conversation_chunks).to eq(3)
crafted_system_prompt = custom_persona.craft_prompt(with_cc).messages.first[:content]
expect(crafted_system_prompt).to include("fragment-n0")
expect(crafted_system_prompt).to include("fragment-n1")
expect(crafted_system_prompt).to include("fragment-n2")
expect(crafted_system_prompt).not_to include("fragment-n3")
end
end
context "when the reranker is available" do
before do
SiteSetting.ai_hugging_face_tei_reranker_endpoint = "https://test.reranker.com"
# hard coded internal implementation, reranker takes x5 number of chunks
stub_fragments(15, expected_limit: 50) # Mimic limit being more than 10 results
end
it "uses the re-ranker to reorder the fragments and pick the top 10 candidates" do
expected_reranked = (0..14).to_a.reverse.map { |idx| { index: idx } }
WebMock.stub_request(:post, "https://test.reranker.com/rerank").to_return(
status: 200,
body: JSON.dump(expected_reranked),
)
crafted_system_prompt = ai_persona.craft_prompt(with_cc).messages.first[:content]
expect(crafted_system_prompt).to include("fragment-n14")
expect(crafted_system_prompt).to include("fragment-n13")
expect(crafted_system_prompt).to include("fragment-n12")
expect(crafted_system_prompt).not_to include("fragment-n4") # Fragment #11 not included
end
end
context "when the reranker is not available" do
before { stub_fragments(10) }
it "picks the first 10 candidates from the similarity search" do
crafted_system_prompt = ai_persona.craft_prompt(with_cc).messages.first[:content]
expect(crafted_system_prompt).to include("fragment-n0")
expect(crafted_system_prompt).to include("fragment-n1")
expect(crafted_system_prompt).to include("fragment-n2")
expect(crafted_system_prompt).not_to include("fragment-n10") # Fragment #10 not included
end
end
end
end
end