Revert the block parts of the compaction view

This reverts commits 4b9b226 and 56129a5. A compaction sends its view
again as one input part, followed by the task and, on a retry, the note.

The ChatGPT backend writes a cache entry only at the end of the latest
message of a request, and a later request reads it only at one of its
own message ends. Parts inside a message are not lookup boundaries, so
the blocks gave no cache hit. The entry of a compaction ends after its
task, which no later compaction begins with, so first tries miss with
any split of the view. A warm-up call per prefix would write a usable
entry, but it costs more than the misses, so the first-try misses stay.

Issue: 8c5fd2a

diff --git a/compactor.go b/compactor.go
index 11497b8ddd6aeb919a75cf37c7ae08e8f2c25c59..16303b0c342cc7a45784754076501e53344918e2 100644 (file)
--- a/compactor.go
+++ b/compactor.go
@@ -4,18 +4,12 @@ import (
        "context"
        "errors"
        "fmt"
        "context"
        "errors"
        "fmt"

-       "slices"

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

-const (
-       // compactionTries is the most calls for one node.
-       compactionTries = 5
-
-       // viewBlock is the number of view lines in one input part.
-       viewBlock = 4
-)
+// 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.
 
 // 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.


@@ -24,10 +18,7 @@ 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
 // 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, in the
-// parts that compactionParts gives. Only the last part holds the end of the
-// view, so compactions whose views end at different lines begin with the same
-// parts up to their last shared block.
+// 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
 //
 // 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


@@ -56,18 +47,18 @@ func newCompactor(client *Client, name string, model ModelConfig) *Compactor {
 
 // Summarize returns the line of the node of task.
 func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
 
 // 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
        }
 
        text, err := compactionTask(task)
        if err != nil {
                return "", err
        }
 

-       first := compactionParts(task.Context, text)

        var lines []string
        note := ""
        for len(lines) < compactionTries {
        var lines []string
        note := ""
        for len(lines) < compactionTries {

-               parts := first
+               parts := []string{view, text}

                if note != "" {
                if note != "" {

-                       parts = append(slices.Clip(first), note)
+                       parts = append(parts, note)

                }
 
                var response Response
                }
 
                var response Response


@@ -104,22 +95,6 @@ func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
        return shortest(lines), nil
 }
 
        return shortest(lines), nil
 }
 

-// compactionParts returns the input parts of a first try for the compaction
-// view context and the task text: the opening tag, one part for each whole
-// block of viewBlock lines, and then one part with the rest of the lines, the
-// closing tag, a newline, and the task. Joined, the parts are the view, a
-// newline, and the task.
-func compactionParts(context, task string) []string {
-       lines := strings.SplitAfter(strings.TrimPrefix(context, "<chat>\n"), "\n")
-       whole := len(lines) - 1 - (len(lines)-1)%viewBlock
-       parts := []string{"<chat>\n"}
-       for n := 0; n < whole; n += viewBlock {
-               parts = append(parts, strings.Join(lines[n:n+viewBlock], ""))
-       }
-
-       return append(parts, strings.Join(lines[whole:], "")+"\n"+task)
-}
-

 // compactionTask returns the task of a build, with its input.
 func compactionTask(task Task) (string, error) {
        if task.Level == 0 && len(task.Input) == 1 {
 // compactionTask returns the task of a build, with its input.
 func compactionTask(task Task) (string, error) {
        if task.Level == 0 && len(task.Input) == 1 {

diff --git a/compactor_test.go b/compactor_test.go
index 20862c77e825022e31d8866da08731e1f5a3473a..35e09eb28b4562530af8dae474fa13d62b49ad8f 100644 (file)
--- a/compactor_test.go
+++ b/compactor_test.go
@@ -20,7 +20,7 @@ func TestCompactorSendsThePromptToolsAndTask(t *testing.T) {
        line, err := compactor.Summarize(t.Context(), Task{
                Level:   0,
                Index:   7,
        line, err := compactor.Summarize(t.Context(), Task{
                Level:   0,
                Index:   7,

-               Context: "<chat>\n0+4|a\n4+1|b\n5+1|c\n6+1|d\n7+1|e\n</chat>",
+               Context: "<chat>\n0+4|user: earlier\n</chat>",

                Input:   []string{"user: " + strings.Repeat("x", 600)},
        })
        require.NoError(t, err)
                Input:   []string{"user: " + strings.Repeat("x", 600)},
        })
        require.NoError(t, err)


@@ -42,66 +42,14 @@ func TestCompactorSendsThePromptToolsAndTask(t *testing.T) {
 
        parts := inputParts(t, server.Call(0))
        require.Equal(t, []string{
 
        parts := inputParts(t, server.Call(0))
        require.Equal(t, []string{

-               "<chat>\n",
-               "0+4|a\n4+1|b\n5+1|c\n6+1|d\n",
-               "7+1|e\n</chat>\n" +
-                       "Compaction: compress message 7 into one line of at most 512 bytes\n" +
+               "<chat>\n0+4|user: earlier\n</chat>\n",
+               "Compaction: compress message 7 into one line of at most 512 bytes\n" +

                        "(about 70 words), the length of this ruler:\n" +
                        strings.Repeat("-", 512) + "\n" +
                        "<input>\nuser: " + strings.Repeat("x", 600) + "\n</input>",
        }, parts)
 }
 
                        "(about 70 words), the length of this ruler:\n" +
                        strings.Repeat("-", 512) + "\n" +
                        "<input>\nuser: " + strings.Repeat("x", 600) + "\n</input>",
        }, parts)
 }
 

-func TestCompactionPartsEndWithTheRestAndTheTask(t *testing.T) {
-       testCases := []struct {
-               name    string
-               context string
-               want    []string
-       }{
-               {
-                       name:    "empty view has the opening tag and the last part",
-                       context: "<chat>\n</chat>",
-                       want:    []string{"<chat>\n", "</chat>\nTASK"},
-               },
-               {
-                       name:    "lines short of a block go into the last part",
-                       context: "<chat>\n0+1|a\n1+1|b\n2+1|c\n</chat>",
-                       want:    []string{"<chat>\n", "0+1|a\n1+1|b\n2+1|c\n</chat>\nTASK"},
-               },
-               {
-                       name:    "one whole block",
-                       context: "<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n</chat>",
-                       want:    []string{"<chat>\n", "0+1|a\n1+1|b\n2+1|c\n3+1|d\n", "</chat>\nTASK"},
-               },
-               {
-                       name:    "two whole blocks and a rest",
-                       context: "<chat>\n0+2|a\n2+1|b\n3+1|c\n4+1|d\n5+1|e\n6+1|f\n7+1|g\n8+1|h\n9+1|i\n</chat>",
-                       want: []string{
-                               "<chat>\n",
-                               "0+2|a\n2+1|b\n3+1|c\n4+1|d\n",
-                               "5+1|e\n6+1|f\n7+1|g\n8+1|h\n",
-                               "9+1|i\n</chat>\nTASK",
-                       },
-               },
-       }
-
-       for _, tc := range testCases {
-               t.Run(tc.name, func(t *testing.T) {
-                       parts := compactionParts(tc.context, "TASK")
-                       require.Equal(t, tc.want, parts)
-                       require.Equal(t, tc.context+"\nTASK", strings.Join(parts, ""))
-               })
-       }
-}
-
-func TestCompactionPartsOfALongerViewBeginWithTheBlocksOfAShorterOne(t *testing.T) {
-       shorter := compactionParts("<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n4+1|e\n</chat>", "task 5")
-       longer := compactionParts("<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n4+1|e\n5+1|f\n6+1|g\n7+1|h\n</chat>", "task 8")
-       require.Len(t, shorter, 3)
-       require.Equal(t, shorter[:2], longer[:2])
-       require.NotEqual(t, shorter[2], longer[2])
-}
-

 func TestCompactorWritesTheMergeTask(t *testing.T) {
        text, err := compactionTask(Task{
                Level: 2,
 func TestCompactorWritesTheMergeTask(t *testing.T) {
        text, err := compactionTask(Task{
                Level: 2,