Add Meta-prompting

CHANGELOG.md

@@ -7,6 +7,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## [Unreleased]
 
 ### Added
-- First iteration of the GUI
 
-### Fixed
+Chat tab
+- added delete icon to the last message in the conversation (for easy deletion)
+- added a button in "Advanced Settings" to "Fork a conversation" (save to history for reference, but continue from a fresh copy)
+- added a focus when a template is selected (on expand, on template selection, etc)
+
+Meta-prompting tab
+- Add an experimental meta-prompting experience based on [arxiv](https://arxiv.org/pdf/2401.12954). See the tab "Meta-Prompting" for more details.
+
+### Fixed
+
+## [0.1.0]
+
+### Added
+- The first iteration of the GUI released

Project.toml

@@ -4,12 +4,14 @@ authors = ["J S <49557684+svilupp@users.noreply.github.com> and contributors"]
 version = "0.2.0-DEV"
 
 [deps]
+Dates = "ade2ca70-3891-5945-98fb-dc099432e06a"
 GenieFramework = "a59fdf5c-6bf0-4f5d-949c-a137c9e2f353"
 GenieSession = "03cc5b98-4f21-4eb6-99f2-22eced81f962"
 PromptingTools = "670122d1-24a8-4d70-bfce-740807c42192"
 
 [compat]
 Aqua = "0.7"
+Dates = "<0.0.1, 1"
 GenieFramework = "2.1"
 GenieSession = "1"
 PromptingTools = "0.23"

app.jl

@@ -41,6 +41,9 @@ const HISTORY_SAVE = get(ENV, "PROTO_HISTORY_SAVE", true)
     @in selected_page = "chat"
     @in ministate = true
     # configuration
+    ## @in chat_tracker_tokens_in = 0
+    ## @in chat_tracker_tokens_out = 0
+    ## @in chat_tracker_cost = 0.0
     @in model = isempty(PT.GROQ_API_KEY) ? "gpt4t" : "gllama370"
     @in model_input = ""
     @in model_submit = false
@@ -80,6 +83,15 @@ const HISTORY_SAVE = get(ENV, "PROTO_HISTORY_SAVE", true)
     @in chat_reset = false
     @in chat_rm_last_msg = false
     @in chat_fork = false
+    # Meta Prompting
+    @in meta_submit = false
+    @in meta_reset = false
+    @in meta_disabled = false
+    @in meta_question = ""
+    @in meta_rounds_max = 5
+    @in meta_rounds_current = 0
+    @in meta_displayed = Dict{Symbol, Any}[]
+    @in meta_rm_last_msg = false
     # Template browser
     @in template_filter = ""
     @in template_submit = false
@@ -118,32 +130,18 @@ const HISTORY_SAVE = get(ENV, "PROTO_HISTORY_SAVE", true)
     end
     @onbutton chat_reset begin
         @info "> Chat Reset!"
-        timestamp = Dates.format(now(), "YYYYmmdd_HHMMSS")
-        name = "Conv. @ $timestamp"
-        ## update the chat with edits by the user
-        conv_rendered = render_messages(conv_displayed, chat_template_variables)
-        # Conv. display is already up-to-date, no need to update it!
-        label = label_conversation(conv_rendered; model = model)
-        if HISTORY_SAVE
-            label_clean = replace(label, r"[:\s\"]+" => "_") |> lowercase
-            ## save to disk + to chat history
-            path = joinpath(
-                HISTORY_DIR, "conversation__$(timestamp)__$(label_clean).json")
-            PT.save_conversation(path, conv_rendered)
-            @info "> Chat saved to $path"
-        end
-        history = push!(history,
-            Dict(:name => name, :label => label, :messages => conv_rendered))
-
+        record = save_conversation(
+            conv_displayed; save = HISTORY_SAVE, save_path = HISTORY_DIR,
+            variables = chat_template_variables, model = model)
+        history = push!(history, record)
         ## clean the chat
         conv_displayed = empty!(conv_displayed)
         chat_template_variables = empty!(chat_template_variables)
         chat_question, chat_auto_template, chat_template_selected = "", "", ""
         chat_disabled, chat_advanced_expanded, chat_template_expanded = false, false, false
         # set defaults again
         chat_code_airetry, chat_code_eval = false, false
-        chat_code_prefix = ""
-        chat_temperature = 0.7
+        chat_code_prefix, chat_temperature = "", 0.7
     end
     @onbutton chat_submit begin
         chat_disabled = true
@@ -233,6 +231,47 @@ const HISTORY_SAVE = get(ENV, "PROTO_HISTORY_SAVE", true)
         chat_reset = true
         conv_displayed = conv_displayed_temp
     end
+    ### Meta-prompting
+    @onbutton meta_submit begin
+        meta_disabled = true
+        if meta_rounds_current < meta_rounds_max
+            # we skip prepare_conversation to avoid create user+system prompt when we start, just grab the messages
+            conv_current = render_messages(meta_displayed)
+            while meta_rounds_current < meta_rounds_max
+                meta_rounds_current = meta_rounds_current + 1
+                ## update conv, but indicate if it's final_answer
+                early_stop, conv_current = meta_prompt_step!(
+                    conv_current; counter = meta_rounds_current, model = model, question = meta_question)
+                meta_displayed = [msg2display(msg; id)
+                                  for (id, msg) in enumerate(conv_current)]
+                early_stop && break
+            end
+        elseif meta_question != ""
+            @info "> Meta-prompting follow up question!"
+            conv = prepare_conversation(meta_displayed; question = meta_question)
+            conv_current = send_to_model(conv; model = model)
+            meta_displayed = [msg2display(msg; id)
+                              for (id, msg) in enumerate(conv_current)]
+        end
+        meta_disabled, meta_question = false, ""
+    end
+    @onbutton meta_reset begin
+        @info "> Meta-Prompting Reset!"
+        record = save_conversation(
+            meta_displayed; save = HISTORY_SAVE, save_path = HISTORY_DIR,
+            model = model, file_prefix = "conversation__meta")
+        history = push!(history, record)
+        ## clean the messages
+        meta_rounds_current = 0
+        meta_displayed = empty!(meta_displayed)
+        meta_disabled, meta_question, meta_rounds_current = false, "", 0
+    end
+    @onbutton meta_rm_last_msg begin
+        @info "> Deleting last turn!"
+        meta_rounds_current = meta_rounds_current - 1
+        pop!(meta_displayed)
+        meta_displayed = meta_displayed
+    end
     ### Template browsing behavior
     @onbutton template_submit begin
         @info "> Template filter: $template_filter"
@@ -296,7 +335,7 @@ end
             this.$refs.tpl_select.focus();
         });
     },
-    filterFn (val, update) {
+    filterFn(val, update) {
         if (val === '') {
             update(() => {
             // reset to full option list
@@ -310,7 +349,7 @@ end
             this.chat_template_options = this.chat_template_options_all.filter(v => v.toLowerCase().indexOf(needle) > -1)
         })
         },
-    filterFnAuto (val, update) {
+    filterFnAuto(val, update) {
         if (val === '') {
             update(() => {
             // reset to full option list
@@ -325,8 +364,7 @@ end
             this.chat_auto_template_options = this.chat_auto_template_options_all.filter(v => v.toLowerCase().indexOf(needle) > -1)
         })
         },
-    copyToClipboard: function(index) {
-        console.log(index);
+    copyToClipboard(index) {
         const str = this.conv_displayed[index].content; // extract the content of the element in position `index`
         const el = document.createElement('textarea');  // Create a <textarea> element
         el.value = str;                                 // Set its value to the string that you want copied
@@ -337,6 +375,18 @@ end
         el.select();                                    // Select the <textarea> content
         document.execCommand('copy');                   // Copy - only works as a result of a user action (e.g. click events)
         document.body.removeChild(el);                  // Remove the <textarea> element
+    },
+    copyToClipboardMeta(index) {
+        const str = this.meta_displayed[index].content; // extract the content of the element in position `index`
+        const el = document.createElement('textarea');  // Create a <textarea> element
+        el.value = str;                                 // Set its value to the string that you want copied
+        el.setAttribute('readonly', '');                // Make it readonly to be tamper-proof
+        el.style.position = 'absolute';                 
+        el.style.left = '-9999px';                      // Move outside the screen to make it invisible
+        document.body.appendChild(el);                  // Append the <textarea> element to the HTML document
+        el.select();                                    // Select the <textarea> content
+        document.execCommand('copy');                   // Copy - only works as a result of a user action (e.g. click events)
+        document.body.removeChild(el);                  // Remove the <textarea> element
     }
     """
 end

src/ProToPortal.jl

@@ -1,9 +1,10 @@
 module ProToPortal
 
+using Dates
 using PromptingTools
 const PT = PromptingTools
 using PromptingTools: JSON3
-using PromptingTools: AICode, last_message, last_output
+using PromptingTools: AICode, last_message, last_output, save_conversation
 using PromptingTools.Experimental.AgentTools: aicodefixer_feedback, airetry!, AICall,
                                               AIGenerate
 const AT = PromptingTools.Experimental.AgentTools
@@ -21,13 +22,17 @@ export render_messages, render_template_messages
 export conversation2transcript, parse_critic, load_conversations_from_dir
 include("utils.jl")
 
+export save_conversation
+include("serialization.jl")
+
 export flash, flash_has_message
 include("flash.jl")
 
 export messagecard, templatecard
 include("components.jl")
 
 include("view_chat.jl")
+include("view_meta.jl")
 
 export ui, ui_login
 include("view.jl")
@@ -39,6 +44,9 @@ export send_to_model, prepare_conversation, label_conversation, evaluate_code
 export build_lazy_aicall, autofix_code
 include("llm.jl")
 
+export meta_prompt_step!
+include("meta_prompting.jl")
+
 function __init__()
     ## Load extra templates
     PT.load_templates!(joinpath(@__DIR__, "..", "templates"); remember_path = true) # add our custom ones

src/llm.jl

@@ -1,7 +1,7 @@
 "Prepares the conversation for sending to the LLM"
 function prepare_conversation(
         display::Vector{Dict{Symbol, Any}}, placeholders::Vector{Dict{
-            Symbol, Any}};
+            Symbol, Any}} = Vector{Dict{Symbol, Any}}();
         question::String = "", template::String = "", system_prompt::String = "")
     conv = render_messages(display, placeholders)
     if template != ""
@@ -28,15 +28,16 @@ function prepare_conversation(
 end
 
 "Code evaluator. Returns the evaluted code block (AICode) and the feedback string."
-function evaluate_code(conv::AbstractVector{<:PT.AbstractMessage}; prefix = "")
-    @info "> Evaluating code"
+function evaluate_code(conv::AbstractVector{<:PT.AbstractMessage};
+        prefix::String = "", header::String = "## Code Evaluation")
+    @info ">> Evaluating code"
     cb = AICode(
         last(conv); prefix, skip_unsafe = true, capture_stdout = true)
-    @info "> Code Success: $(isvalid(cb))"
+    @info ">> Code Success: $(isvalid(cb))"
 
     ### Build the response
     io = IOBuffer()
-    println(io, "## Code Evaluation", "\n")
+    println(io, header, "\n")
     if isvalid(cb)
         println(io, "**Outcome:** Code is valid", "\n")
         println(io, "**Output:**\n $(cb.stdout)", "\n")
@@ -50,6 +51,13 @@ function evaluate_code(conv::AbstractVector{<:PT.AbstractMessage}; prefix = "")
     return cb, feedback
 end
 
+# Convenience function for evaluating code from string
+# AIMessage is the better input because it handles code extraction with a few fallbacks!
+function evaluate_code(str::AbstractString;
+        prefix::String = "", header::String = "## Code Evaluation")
+    evaluate_code([PT.AIMessage(str)]; prefix, header)
+end
+
 "Constructs AIGenerate call that mimics if it was just executed"
 function build_lazy_aicall(conv::AbstractVector{<:PT.AbstractMessage};
         model::String, max_retries::Int = 3, n_samples::Int = 2)
@@ -83,6 +91,14 @@ function send_to_model(
     result = aigenerate(conv; model, api_kwargs = (; temperature), return_all = true)
     return result
 end
+"Sends the conversation to the LLM."
+function send_to_model(
+        any_template::Symbol; model::String, temperature::Float64 = 0.7, kwargs...)
+    @info ">> Sending AITemplate $(any_template) to LLM with temp $temperature"
+    result = aigenerate(
+        any_template; model, api_kwargs = (; temperature), return_all = true, kwargs...)
+    return result
+end
 "Sends the conversation to the Auto-Critic Template for evaluation and suggestions"
 function send_to_model(critic_template::Symbol,
         conv::AbstractVector{<:PT.AbstractMessage}; model::String, temperature::Float64 = 0.7)