Build the summary nodes with a model

[onechat] / compactor.go

package main

import (
        "context"
        "errors"
        "fmt"
        "strings"
        "unicode/utf8"
)

// compactionTries is the most calls for one node.
const compactionTries = 5

// ruler shows the model the size limit of a line. Models cannot count bytes,
// and a sample line as the ruler gets its content copied.
var ruler = strings.Repeat("-", nodeBytes)

// Compactor builds the nodes of one agent with a model. A compaction is a call
// like a turn, with the same system prompt and tools, so it reads them from the
// prompt cache of the turns. The tools are never called. After them comes one
// user message with the compaction view up to the node and the task.
//
// When the line is over nodeBytes, the next try sends the same view and task,
// plus a note with the size of the newest line and its cut at the limit. A try
// never sends the model's earlier reply or reasoning back, so every try has the
// same cached prefix and about the same size. After at most compactionTries
// tries, the shortest line is the node. A few bytes over the limit is fine,
// because the view measures the real sizes. A call that fails before any line
// fails the build, and the core tries it again at the next message.
type Compactor struct {
        client       *Client
        model        ModelConfig
        instructions string
        tools        []Tool
}

// newCompactor returns the compactor of the agent name, which sends its calls
// through client with model.
func newCompactor(client *Client, name string, model ModelConfig) *Compactor {
        return &Compactor{
                client:       client,
                model:        model,
                instructions: prompt(name),
                tools:        agentTools(),
        }
}

// Summarize returns the line of the node of task.
func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
        view := task.Context + "\n"
        text, err := compactionTask(task)
        if err != nil {
                return "", err
        }

        var lines []string
        note := ""
        for len(lines) < compactionTries {
                parts := []string{view, text}
                if note != "" {
                        parts = append(parts, note)
                }

                var response Response
                response, err = c.client.Create(ctx, Request{
                        Model:        c.model.Model,
                        Effort:       c.model.Effort,
                        Instructions: c.instructions,
                        Tools:        c.tools,
                        Input:        []any{userMessage(parts...)},
                })
                if err != nil {
                        break
                }

                line := strings.TrimSpace(response.Text)
                if line == "" {
                        err = errors.New("compaction gave no line")
                        break
                }

                lines = append(lines, line)
                if len(line) <= nodeBytes {
                        break
                }

                note = tooLong(line)
        }

        if len(lines) == 0 {
                return "", fmt.Errorf("compact node (%d, %d): %w", task.Level, task.Index, err)
        }

        return shortest(lines), nil
}

// compactionTask returns the task of a build, with its input.
func compactionTask(task Task) (string, error) {
        if task.Level == 0 && len(task.Input) == 1 {
                return fmt.Sprintf(`Compaction: compress message %d into one line of at most %d bytes
(about 70 words), the length of this ruler:
%s
<input>
%s
</input>`, task.Index, nodeBytes, ruler, task.Input[0]), nil
        }

        if task.Level > 0 && len(task.Input) == 2 {
                first := task.Index << task.Level
                half := 1 << (task.Level - 1)

                return fmt.Sprintf(`Compaction: merge lines %d+%d and %d+%d, adjacent, into one line of at most
%d bytes (about 70 words), the length of this ruler:
%s
<chat> may hold their messages, %d to %d, in more detail: take details
of them from there too.
<input>
%s
%s
</input>`, first, half, first+half, half, nodeBytes, ruler, first, nodeEnd(task.Level, task.Index)-1, task.Input[0], task.Input[1]), nil
        }

        return "", fmt.Errorf("node (%d, %d) has %d inputs", task.Level, task.Index, len(task.Input))
}

// tooLong returns the note for a line over nodeBytes: its size and its first
// nodeBytes bytes, which show the model where the limit cuts it.
func tooLong(line string) string {
        return fmt.Sprintf(`Too long: your last line for this <input> was %d bytes,
over the %d-byte limit. Write the whole line again, cutting just
enough of the least valuable items to fit before this cut:
%s| ← LIMIT`, len(line), nodeBytes, cutLine(line))
}

// cutLine returns the first nodeBytes bytes of line, without a part of a
// character at the end.
func cutLine(line string) string {
        if len(line) <= nodeBytes {
                return line
        }

        cut := line[:nodeBytes]
        for !utf8.ValidString(cut) {
                cut = cut[:len(cut)-1]
        }

        return cut
}

// shortest returns the first of the shortest lines.
func shortest(lines []string) string {
        best := lines[0]
        for _, line := range lines[1:] {
                if len(line) < len(best) {
                        best = line
                }
        }

        return best
}