flightlesssomething/openai.go

package flightlesssomething

import (
	"context"
	"errors"
	"log"
	"math"
	"sort"
	"strconv"
	"strings"
	"sync"

	anthro "github.com/liushuangls/go-anthropic"
	openai "github.com/sashabaranov/go-openai"
)

const systemMessage = `
You are given PC benchmark data of several runs. All this data is visible in the website in a form of charts and your goal is to provide insights.

Required format MUST be like this:
# Top runs:
* **Highest FPS**: Run 1 and comment
* **Smoothest FPS**: Run 2 and comment
* **Best overall**: Run 3 and comment
# Issues:
* ...
* ...
* ...
# Summary
...

You MUST:
* "Top runs" section is needed ONLY if there are 2 runs or more.
* In "Top runs" sections, label names must use in-line code syntax.
* In Issues section, Figure out if any of the run is significantly worse then others in the same benchmark. You MUST use ONLY the data provided to explain the difference, and your points must be based only on the data provided. If there are no issues - do NOT write this section. Do not make any guesses. Additional requirements: (a) validate if the same hardware/software was used (by using provided text fields, NOT the data), (b) do not speculate, but use numbers to back up your claims, (c) only write if it's an actual issue with FPS (everything else is just additional information).
* In Top runs section, provide which run has the (average) "Highest FPS", which has the "Smoothest FPS" (LOWEST std.dev. and variance of FPS value - LOWEST, NOT HIGHEST) and which is the best "Best overall" (preferrably lower std.dev./variance than higher FPS, but if slight decrease in stability gives significantly higher FPS - pick that one). NEVER consider runs that have significantly lower FPS or has other significant issues. Exclude runs from consideration if they are significantly worse than the rest (as it would be mentioned in issues section). Note that your goal is to pick winners and not do a comparison in this section. Include numbers to justify your claims.
* In Summary section, provide an overview/summary of this benchmark in only one paragraph and in natural language. In this section, use your deep understanding of context of what is actually being compared (driver versions? Different configurations? Different settings? different schedulers? Different OS? Focus on that) and comment on that.
* First 2 sections should be bullet points, no subpoints, only 1 bullet point per point, while summary should be a single paragraph.
* NEVER use actual numbers. Instead, use percentage in comparison to other runs.

Your data does not contain anything about scx schedulers, but some benchmarks might compare these:
* sched_ext is a Linux kernel feature which enables implementing kernel thread schedulers in BPF and dynamically loading them. This repository contains various scheduler implementations and support utilities.
* sched_ext enables safe and rapid iterations of scheduler implementations, thus radically widening the scope of scheduling strategies that can be experimented with and deployed; even in massive and complex production environments.
* Available scx schedulers (at the time of writting): scx_asdf scx_bpfland scx_central scx_lavd scx_layered scx_nest scx_qmap scx_rlfifo scx_rustland scx_rusty scx_simple scx_userland

Additional RULES THAT YOU MUST STRICTLY FOLLOW:
1. In summary and issues sections, you are STRICTLY FORBIDDEN TO REFER TO ANY RUN BY IT'S LABEL NAME. Instead, you must use context and deep understand what what is being compared, and refer to each run by it's context, rather than title.
2. In summary section, any technical terms or codes are easier to read when encapsulated with in-code syntax.
3. In summary section, never refer to runs as "first run" or "second run". Refer to them by their context.
4. In issues section, NEVER write "has no issues" kind of statements.
5. In issues section, NEVER write anything that is not DIRECTLY related to issue with either FPS or Frametime.
6. In issues section, additionally point out if configuration is different that isn't directly part of benchmark comparison (e.g. comparing schedulers, but cpu or OS is different)
`

var (
	inProgressSummaries    = map[uint]struct{}{}
	inProgressSummariesMux = &sync.Mutex{}
)

func generateSummary(b *Benchmark, bds []*BenchmarkData) {
	// Check if OpenAI integration is not enabled
	if openaiClient == nil && anthroClient == nil {
		return
	}

	// Lock mutex, as integration is enabled and might be already in progress
	inProgressSummariesMux.Lock()

	// Check if generation is already in progress
	if _, ok := inProgressSummaries[b.ID]; ok {
		inProgressSummariesMux.Unlock()
		return
	}
	inProgressSummaries[b.ID] = struct{}{}
	inProgressSummariesMux.Unlock()

	// Create user prompt
	userPrompt := writeAIPrompt(bds, b.Title, b.Description)

	// Retrieve AI response
	resp, err := getAIResponse(systemMessage, userPrompt)
	if err != nil {
		log.Println("Failed to generate AI summary:", err)
		return
	}

	db.Model(&Benchmark{}).Where("id = ?", b.ID).Update("AiSummary", resp)

	// Update status
	inProgressSummariesMux.Lock()
	delete(inProgressSummaries, b.ID)
	inProgressSummariesMux.Unlock()
}

func getAIResponse(systemMessage string, userPrompt string) (string, error) {
	if openaiClient != nil {
		resp, err := openaiClient.CreateChatCompletion(
			context.Background(),
			openai.ChatCompletionRequest{
				Model:       openaiModel,
				Temperature: 0.0,
				Messages: []openai.ChatCompletionMessage{
					{Role: openai.ChatMessageRoleSystem, Content: systemMessage},
					{Role: openai.ChatMessageRoleUser, Content: userPrompt},
				},
			},
		)
		if err != nil {
			return "", err
		}
		return resp.Choices[0].Message.Content, nil
	} else if anthroClient != nil {
		resp, err := anthroClient.CreateMessages(context.Background(), anthro.MessagesRequest{
			Model:     anthroModel,
			System:    systemMessage,
			MaxTokens: 1024,
			Messages: []anthro.Message{
				anthro.NewUserTextMessage(userPrompt),
			},
		})
		if err != nil {
			return "", err
		}
		return resp.Content[0].Text, nil
	}
	return "", errors.New("no AI client configured")
}

func writeAIPrompt(bds []*BenchmarkData, bdTitle, bdDescription string) string {
	sb := strings.Builder{}
	sb.WriteString("Benchmark title: ")
	sb.WriteString(bdTitle)
	sb.WriteString("\n")
	sb.WriteString("Benchmark description: \n")
	sb.WriteString(bdDescription)
	sb.WriteString("\n\n")
	sb.WriteString("Benchmark contains ")
	sb.WriteString(strconv.Itoa(len(bds)))
	sb.WriteString(" runs:\n")

	for _, benchmarkRun := range bds {
		sb.WriteString("\nLabel: ")
		sb.WriteString(benchmarkRun.Label)
		sb.WriteString("\n")

		sb.WriteString("OS: ")
		sb.WriteString(benchmarkRun.SpecOS)
		sb.WriteString("\n")
		sb.WriteString("GPU: ")
		sb.WriteString(benchmarkRun.SpecGPU)
		sb.WriteString("\n")
		sb.WriteString("CPU: ")
		sb.WriteString(benchmarkRun.SpecCPU)
		sb.WriteString("\n")
		sb.WriteString("RAM: ")
		sb.WriteString(benchmarkRun.SpecRAM)
		sb.WriteString("\n")
		sb.WriteString("Linux kernel: ")
		sb.WriteString(benchmarkRun.SpecLinuxKernel)
		sb.WriteString("\n")
		sb.WriteString("Linux scheduler: ")
		sb.WriteString(benchmarkRun.SpecLinuxScheduler)
		sb.WriteString("\n")

		// FPS
		stats := calculateAIPromptArrayStats(benchmarkRun.DataFPS)
		sb.WriteString("FPS: ")
		sb.WriteString(stats.String())
		sb.WriteString("\n")

		// Frame time
		stats = calculateAIPromptArrayStats(benchmarkRun.DataFrameTime)
		sb.WriteString("Frame time: ")
		sb.WriteString(stats.String())
		sb.WriteString("\n")

		// CPU load
		stats = calculateAIPromptArrayStats(benchmarkRun.DataCPULoad)
		sb.WriteString("CPU load: ")
		sb.WriteString(stats.String())
		sb.WriteString("\n")

		// GPU load
		stats = calculateAIPromptArrayStats(benchmarkRun.DataGPULoad)
		sb.WriteString("GPU load: ")
		sb.WriteString(stats.String())
		sb.WriteString("\n")

		// CPU temp
		stats = calculateAIPromptArrayStats(benchmarkRun.DataCPUTemp)
		sb.WriteString("CPU temp: ")
		sb.WriteString(stats.String())
		sb.WriteString("\n")

		// GPU temp
		stats = calculateAIPromptArrayStats(benchmarkRun.DataGPUTemp)
		sb.WriteString("GPU temp: ")
		sb.WriteString(stats.String())
		sb.WriteString("\n")

		// GPU core clock
		stats = calculateAIPromptArrayStats(benchmarkRun.DataGPUCoreClock)
		sb.WriteString("GPU core clock: ")
		sb.WriteString(stats.String())
		sb.WriteString("\n")

		// GPU mem clock
		stats = calculateAIPromptArrayStats(benchmarkRun.DataGPUMemClock)
		sb.WriteString("GPU mem clock: ")
		sb.WriteString(stats.String())
		sb.WriteString("\n")

		// GPU VRAM used
		stats = calculateAIPromptArrayStats(benchmarkRun.DataGPUVRAMUsed)
		sb.WriteString("GPU VRAM used: ")
		sb.WriteString(stats.String())
		sb.WriteString("\n")

		// GPU power
		stats = calculateAIPromptArrayStats(benchmarkRun.DataGPUPower)
		sb.WriteString("GPU power: ")
		sb.WriteString(stats.String())
		sb.WriteString("\n")

		// RAM used
		stats = calculateAIPromptArrayStats(benchmarkRun.DataRAMUsed)
		sb.WriteString("RAM used: ")
		sb.WriteString(stats.String())
		sb.WriteString("\n")

		// Swap used
		stats = calculateAIPromptArrayStats(benchmarkRun.DataSwapUsed)
		sb.WriteString("Swap used: ")
		sb.WriteString(stats.String())
		sb.WriteString("\n")
	}

	return sb.String()
}

type AIPromptArrayStats struct {
	Count        int
	Low1Percent  float64
	Mean         float64
	Median       float64
	Top97Percent float64
	StdDev       float64
	Variance     float64
}

func calculateAIPromptArrayStats(data []float64) AIPromptArrayStats {
	if len(data) == 0 {
		return AIPromptArrayStats{}
	}

	sort.Float64s(data)
	count := len(data)

	low1PercentIndex := int(math.Ceil(0.01*float64(count))) - 1
	if low1PercentIndex < 0 {
		low1PercentIndex = 0
	}
	low1Percent := data[low1PercentIndex]

	top97PercentIndex := int(math.Ceil(0.97*float64(count))) - 1
	if top97PercentIndex < 0 {
		top97PercentIndex = 0
	}
	top97Percent := data[top97PercentIndex]

	mean := 0.0
	for _, value := range data {
		mean += value
	}
	mean /= float64(count)

	median := 0.0
	if count%2 == 0 {
		median = (data[count/2-1] + data[count/2]) / 2
	} else {
		median = data[count/2]
	}

	variance := 0.0
	for _, value := range data {
		variance += (value - mean) * (value - mean)
	}
	variance /= float64(count)

	stdDev := math.Sqrt(variance)

	return AIPromptArrayStats{
		Count:        count,
		Low1Percent:  low1Percent,
		Mean:         mean,
		Median:       median,
		Top97Percent: top97Percent,
		StdDev:       stdDev,
		Variance:     variance,
	}
}

func (as AIPromptArrayStats) String() string {
	return strings.Join([]string{
		"Count:" + strconv.Itoa(as.Count),
		"Low1Percent:" + strconv.FormatFloat(as.Low1Percent, 'f', -1, 64),
		"Mean:" + strconv.FormatFloat(as.Mean, 'f', -1, 64),
		"Median:" + strconv.FormatFloat(as.Median, 'f', -1, 64),
		"Top97Percent:" + strconv.FormatFloat(as.Top97Percent, 'f', -1, 64),
		"StdDev:" + strconv.FormatFloat(as.StdDev, 'f', -1, 64),
		"Variance:" + strconv.FormatFloat(as.Variance, 'f', -1, 64),
	}, ", ")
}