git.giacomolli.uk - onechat/blob - compactor.go

   1 package main
   2
   3 import (
   4         "context"
   5         "errors"
   6         "fmt"
   7         "strings"
   8         "unicode/utf8"
   9 )
  10
  11 // compactionTries is the most calls for one node.
  12 const compactionTries = 5
  13
  14 // ruler shows the model the size limit of a line. Models cannot count bytes,
  15 // and a sample line as the ruler gets its content copied.
  16 var ruler = strings.Repeat("-", nodeBytes)
  17
  18 // Compactor builds the nodes of one agent with a model. A compaction is a call
  19 // like a turn, with the same system prompt and tools, so it reads them from the
  20 // prompt cache of the turns. The tools are never called. After them comes one
  21 // user message with the compaction view up to the node and the task.
  22 //
  23 // When the line is over nodeBytes, the next try sends the same view and task,
  24 // plus a note with the size of the newest line and its cut at the limit. A try
  25 // never sends the model's earlier reply or reasoning back, so every try has the
  26 // same cached prefix and about the same size. After at most compactionTries
  27 // tries, the shortest line is the node. A few bytes over the limit is fine,
  28 // because the view measures the real sizes. A call that fails before any line
  29 // fails the build, and the core tries it again at the next message.
  30 type Compactor struct {
  31         client       *Client
  32         model        ModelConfig
  33         instructions string
  34         tools        []Tool
  35 }
  36
  37 // newCompactor returns the compactor of the agent name, which sends its calls
  38 // through client with model.
  39 func newCompactor(client *Client, name string, model ModelConfig) *Compactor {
  40         return &Compactor{
  41                 client:       client,
  42                 model:        model,
  43                 instructions: prompt(name),
  44                 tools:        agentTools(),
  45         }
  46 }
  47
  48 // Summarize returns the line of the node of task.
  49 func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
  50         view := task.Context + "\n"
  51         text, err := compactionTask(task)
  52         if err != nil {
  53                 return "", err
  54         }
  55
  56         var lines []string
  57         note := ""
  58         for len(lines) < compactionTries {
  59                 parts := []string{view, text}
  60                 if note != "" {
  61                         parts = append(parts, note)
  62                 }
  63
  64                 var response Response
  65                 response, err = c.client.Create(ctx, Request{
  66                         Model:        c.model.Model,
  67                         Effort:       c.model.Effort,
  68                         Instructions: c.instructions,
  69                         Tools:        c.tools,
  70                         Input:        []any{userMessage(parts...)},
  71                 })
  72                 if err != nil {
  73                         break
  74                 }
  75
  76                 line := strings.TrimSpace(response.Text)
  77                 if line == "" {
  78                         err = errors.New("compaction gave no line")
  79                         break
  80                 }
  81
  82                 lines = append(lines, line)
  83                 if len(line) <= nodeBytes {
  84                         break
  85                 }
  86
  87                 note = tooLong(line)
  88         }
  89
  90         if len(lines) == 0 {
  91                 return "", fmt.Errorf("compact node (%d, %d): %w", task.Level, task.Index, err)
  92         }
  93
  94         return shortest(lines), nil
  95 }
  96
  97 // compactionTask returns the task of a build, with its input.
  98 func compactionTask(task Task) (string, error) {
  99         if task.Level == 0 && len(task.Input) == 1 {
 100                 return fmt.Sprintf(`Compaction: compress message %d into one line of at most %d bytes
 101 (about 70 words), the length of this ruler:
 102 %s
 103 <input>
 104 %s
 105 </input>`, task.Index, nodeBytes, ruler, task.Input[0]), nil
 106         }
 107
 108         if task.Level > 0 && len(task.Input) == 2 {
 109                 first := task.Index << task.Level
 110                 half := 1 << (task.Level - 1)
 111
 112                 return fmt.Sprintf(`Compaction: merge lines %d+%d and %d+%d, adjacent, into one line of at most
 113 %d bytes (about 70 words), the length of this ruler:
 114 %s
 115 <chat> may hold their messages, %d to %d, in more detail: take details
 116 of them from there too.
 117 <input>
 118 %s
 119 %s
 120 </input>`, first, half, first+half, half, nodeBytes, ruler, first, nodeEnd(task.Level, task.Index)-1, task.Input[0], task.Input[1]), nil
 121         }
 122
 123         return "", fmt.Errorf("node (%d, %d) has %d inputs", task.Level, task.Index, len(task.Input))
 124 }
 125
 126 // tooLong returns the note for a line over nodeBytes: its size and its first
 127 // nodeBytes bytes, which show the model where the limit cuts it.
 128 func tooLong(line string) string {
 129         return fmt.Sprintf(`Too long: your last line for this <input> was %d bytes,
 130 over the %d-byte limit. Write the whole line again, cutting just
 131 enough of the least valuable items to fit before this cut:
 132 %s| ← LIMIT`, len(line), nodeBytes, cutLine(line))
 133 }
 134
 135 // cutLine returns the first nodeBytes bytes of line, without a part of a
 136 // character at the end.
 137 func cutLine(line string) string {
 138         if len(line) <= nodeBytes {
 139                 return line
 140         }
 141
 142         cut := line[:nodeBytes]
 143         for !utf8.ValidString(cut) {
 144                 cut = cut[:len(cut)-1]
 145         }
 146
 147         return cut
 148 }
 149
 150 // shortest returns the first of the shortest lines.
 151 func shortest(lines []string) string {
 152         best := lines[0]
 153         for _, line := range lines[1:] {
 154                 if len(line) < len(best) {
 155                         best = line
 156                 }
 157         }
 158
 159         return best
 160 }