// 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)
// 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
// 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
+// 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
return "", err
}
- var first []any
-
- for _, message := range compactorTexts(task.Context, text) {
- first = append(first, modelUserMessage(message))
- }
+ first := []any{modelUserMessage(task.Context + "\n" + text)}
var lines []string
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 {
}, texts)
}
-func TestCompactorSendsEachWholeBlockOfTheViewAsAMessage(t *testing.T) {
- testCases := []struct {
- name string
- context string
- want []string
- }{
- {
- name: "empty view goes with the task",
- context: "<chat>\n</chat>",
- want: []string{"<chat>\n</chat>\nTASK"},
- },
- {
- name: "lines short of a block go with the task",
- context: "<chat>\n0+1|a\n1+1|b\n2+1|c\n</chat>",
- want: []string{"<chat>\n0+1|a\n1+1|b\n2+1|c\n</chat>\nTASK"},
- },
- {
- name: "one whole block and the task",
- context: "<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n</chat>",
- want: []string{"<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n", "</chat>\nTASK"},
- },
- {
- name: "two whole blocks and the rest with the task",
- 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>\n0+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) {
- server := newModelFake(t)
- task := compactorLeafTask()
- task.Context = tc.context
-
- _, err := newTestCompactor(t, server).Summarize(t.Context(), task)
- require.NoError(t, err)
-
- text, err := compactorTask(task)
- require.NoError(t, err)
-
- want := make([]string, len(tc.want))
- for n, message := range tc.want {
- want[n] = strings.Replace(message, "TASK", text, 1)
- }
-
- texts := compactorInputTexts(t, server.Call(0))
- require.Equal(t, want, texts)
- require.Equal(t, tc.context+"\n"+text, strings.Join(texts, ""))
- })
- }
-}
-
-func TestCompactorSendsTheSameBlocksForViewsWithTheSameFirstLines(t *testing.T) {
- shared := "<chat>\n0+4|a\n4+1|b\n5+1|c\n6+1|d\n7+1|e\n8+1|f\n9+1|g\n10+1|h\n"
- server := newModelFake(t)
- compactor := newTestCompactor(t, server)
- _, err := compactor.Summarize(t.Context(), Task{
- Level: 0,
- Index: 11,
- Context: shared + "11+1|i\n</chat>",
- Input: []string{"user: eleven"},
- })
- require.NoError(t, err)
-
- _, err = compactor.Summarize(t.Context(), Task{
- Level: 1,
- Index: 5,
- Context: shared + "11+1|j\n12+1|k\n</chat>",
- Input: []string{"left line", "right line"},
- })
- require.NoError(t, err)
-
- first := compactorInputTexts(t, server.Call(0))
- second := compactorInputTexts(t, server.Call(1))
- require.Len(t, first, 3)
- require.Len(t, second, 3)
- require.Equal(t, first[:2], second[:2])
- require.NotEqual(t, first[2], second[2])
-}
-
func TestCompactorWritesTheMergeTask(t *testing.T) {
text, err := compactorTask(Task{
Level: 2,
require.Equal(t, len(long[n]), record.TextBytes)
}
+ text, err := compactorTask(task)
+ require.NoError(t, err)
+
first := compactorInputTexts(t, server.Call(0))
- require.Len(t, first, 2)
+ require.Equal(t, []string{task.Context + "\n" + text}, first)
for n, want := range []string{long[0], long[1]} {
call := server.Call(n + 1)
texts := compactorInputTexts(t, call)
- require.Equal(t, first, texts[:2])
- require.Equal(t, []string{compactorTooLong(want)}, texts[2:])
- require.Contains(t, texts[2], "was "+strconv.Itoa(len(want))+" bytes,")
- require.Contains(t, texts[2], "\n"+want[:512]+"| <- LIMIT")
+ require.Equal(t, first, texts[:1])
+ require.Equal(t, []string{compactorTooLong(want)}, texts[1:])
+ require.Contains(t, texts[1], "was "+strconv.Itoa(len(want))+" bytes,")
+ require.Contains(t, texts[1], "\n"+want[:512]+"| <- LIMIT")
- note, err := json.Marshal(modelUserMessage(texts[2]))
+ note, err := json.Marshal(modelUserMessage(texts[1]))
require.NoError(t, err)
require.LessOrEqual(t, len(call.raw), len(server.Call(0).raw)+len(",")+len(note))
}
projects / onechat / commitdiff