Carry new images through model request chains

[onechat] / compactor.go

diff --git a/compactor.go b/compactor.go
index be17c15cd43e29b98e433ed4955acbd9cfa22fe3..2b239015ce763f56d1a23a4621647f056023b07b 100644 (file)
--- a/compactor.go
+++ b/compactor.go
@@ -4,7 +4,6 @@ import (
        "context"
        "errors"
        "fmt"
        "context"
        "errors"
        "fmt"

-       "slices"

        "strings"
        "unicode/utf8"
 )
        "strings"
        "unicode/utf8"
 )


@@ -12,48 +11,45 @@ import (
 // compactorTries is the most calls for one node.
 const compactorTries = 5
 
 // compactorTries is the most calls for one node.
 const compactorTries = 5
 

-// compactorBlock is the number of view lines in one block message.
-const compactorBlock = 4
-

 // compactorRuler 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 compactorRuler = strings.Repeat("-", nodeBytes)
 
 // newCompactor returns the compactor of an agent, which sends its calls through
 // client with model. instructions gives the system prompt of the agent at each
 // compactorRuler 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 compactorRuler = strings.Repeat("-", nodeBytes)
 
 // newCompactor returns the compactor of an agent, which sends its calls through
 // client with model. instructions gives the system prompt of the agent at each

-// build, so a build sends the same prompt as the turns of that time.
-func newCompactor(client *ModelClient, instructions func() (string, error), model ModelConfig) *Compactor {
+// build. tools are the tools of that agent, so a build sends the same prompt
+// and tools as its turns.
+func newCompactor(client *ModelClient, instructions func() (string, error), model ModelConfig, tools []ModelTool, images *ImageStore) *Compactor {

        return &Compactor{
                client:       client,
                model:        model,
                instructions: instructions,
        return &Compactor{
                client:       client,
                model:        model,
                instructions: instructions,

-               tools:        promptTools(),
+               tools:        tools,
+               images:       images,

        }
 }
 
 // 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
        }
 }
 
 // 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 come the
-// compaction view up to the node and the task, as the user messages that
-// compactorTexts gives. Each whole block of compactorBlock view lines is its
-// own message, and the last message holds the rest of the lines and the task.
-// The backend reads a cache entry only at a message end of a request, so a
-// compaction whose view begins with the same lines as an earlier one begins
-// with the same block messages.
+// prompt cache of the turns. The tools are never called. After them comes one
+// user message with the compaction view up to the node, its closing tag, a
+// newline, and the task.

 //
 //

-// When the line is over nodeBytes, the next try sends the same messages, plus
+// When the line is over nodeBytes, the next try sends the same message, plus

 // one more user message with a note with the size of the newest line and its
 // one more user message with 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 begins with the whole earlier try and has about the same
-// size. After at most compactorTries 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.
+// cut at the limit. Only the first call of an image leaf sends its pixels. A
+// try never sends the model's earlier reply or reasoning back, so every try
+// begins with the same view and task and has about the same size. After at most
+// compactorTries 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       *ModelClient
        model        ModelConfig
        instructions func() (string, error)
        tools        []ModelTool
 type Compactor struct {
        client       *ModelClient
        model        ModelConfig
        instructions func() (string, error)
        tools        []ModelTool

+       images       *ImageStore

 }
 
 // Summarize returns the line of the node of task.
 }
 
 // Summarize returns the line of the node of task.


@@ -68,19 +64,21 @@ func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
                return "", err
        }
 
                return "", err
        }
 

-       var first []any
-
-       for _, message := range compactorTexts(task.Context, text) {
-               first = append(first, modelUserMessage(message))
+       source := task.Context + "\n" + text
+       parts, err := imageParts(c.images, source, task.Images)
+       if err != nil {
+               return "", fmt.Errorf("read compaction images: %w", err)

        }
 
        }
 

+       first := []any{modelUserParts(parts)}
+

        var lines []string
 
        note := ""
        for len(lines) < compactorTries {
                input := first
                if note != "" {
        var lines []string
 
        note := ""
        for len(lines) < compactorTries {
                input := first
                if note != "" {

-                       input = append(slices.Clip(first), modelUserMessage(note))
+                       input = []any{modelUserMessage(source), modelUserMessage(note)}

                }
 
                var response ModelResponse
                }
 
                var response ModelResponse


@@ -121,12 +119,17 @@ func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
 // compactorTask returns the task of a build, with its input.
 func compactorTask(task Task) (string, error) {
        if task.Level == 0 && len(task.Input) == 1 {
 // compactorTask returns the task of a build, with its input.
 func compactorTask(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
+               text := 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
 (about 70 words), the length of this ruler:
 %s
 <input>
 %s

-</input>`, task.Index, nodeBytes, compactorRuler, task.Input[0]), nil
+</input>`, task.Index, nodeBytes, compactorRuler, task.Input[0])
+               if len(task.Images) > 0 {
+                       text += "\nKeep each [image FILE] name and describe its kind, visible text, amounts, names, and context."
+               }
+
+               return text, nil

        }
 
        if task.Level > 0 && len(task.Input) == 2 {
        }
 
        if task.Level > 0 && len(task.Input) == 2 {


@@ -147,27 +150,6 @@ of them from there too.
        return "", fmt.Errorf("node (%d, %d) has %d inputs", task.Level, task.Index, len(task.Input))
 }
 
        return "", fmt.Errorf("node (%d, %d) has %d inputs", task.Level, task.Index, len(task.Input))
 }
 

-// compactorTexts returns the texts of the user messages of a first try for the
-// compaction view and the task text. Each whole block of compactorBlock lines,
-// counted from the first line of the view, is one message. The last message
-// holds the rest of the lines, the closing tag, a newline, and the task. The
-// first message starts with the opening tag. Joined, the texts are the view, a
-// newline, and the task.
-func compactorTexts(view, task string) []string {
-       body := strings.TrimSuffix(strings.TrimPrefix(view, "<chat>\n"), "</chat>")
-       lines := strings.SplitAfter(body, "\n")
-       lines = lines[:len(lines)-1]
-       whole := len(lines) - len(lines)%compactorBlock
-
-       var texts []string
-
-       for n := 0; n < whole; n += compactorBlock {
-               texts = append(texts, viewMessage(n == 0, lines[n:n+compactorBlock]))
-       }
-
-       return append(texts, viewMessage(whole == 0, lines[whole:])+"</chat>\n"+task)
-}
-

 // compactorTooLong 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 compactorTooLong(line string) string {
 // compactorTooLong 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 compactorTooLong(line string) string {