"context"
"errors"
"fmt"
+ "slices"
"strings"
"unicode/utf8"
)
// 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 comes one
-// user message with the compaction view up to the node and the task.
+// 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.
//
-// 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 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.
+// When the line is over nodeBytes, the next try sends the same messages, 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
+// 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
// 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 := compactorTask(task)
if err != nil {
return "", err
return "", fmt.Errorf("compact node (%d, %d): %w", task.Level, task.Index, err)
}
+ var first []any
+ for _, message := range compactorTexts(task.Context, text) {
+ first = append(first, modelUserMessage(message))
+ }
+
var lines []string
note := ""
for len(lines) < compactorTries {
- parts := []string{view, text}
+ input := first
if note != "" {
- parts = append(parts, note)
+ input = append(slices.Clip(first), modelUserMessage(note))
}
var response ModelResponse
Effort: c.model.Effort,
Instructions: instructions,
Tools: c.tools,
- Input: []any{modelUserMessage(parts...)},
+ Input: input,
Node: &NodeKey{task.Level, task.Index},
})
if err != nil {
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 {
require.NoError(t, err)
require.JSONEq(t, string(tools), string(sent))
- parts := compactorInputParts(t, server.Call(0))
+ texts := compactorInputTexts(t, server.Call(0))
require.Equal(t, []string{
- "<chat>\n0+4|user: earlier\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)
+ }, 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) {
modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem(long[call.number])}, 100, 0, 5))
}
+ task := compactorLeafTask()
+ task.Context = "<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n4+1|e\n</chat>"
+
compactor := newTestCompactor(t, server)
- line, err := compactor.Summarize(t.Context(), compactorLeafTask())
+ line, err := compactor.Summarize(t.Context(), task)
require.NoError(t, err)
require.Equal(t, long[2], line)
require.Equal(t, 3, server.Calls())
require.Equal(t, len(long[n]), record.TextBytes)
}
- first := compactorInputParts(t, server.Call(0))
+ first := compactorInputTexts(t, server.Call(0))
require.Len(t, first, 2)
for n, want := range []string{long[0], long[1]} {
call := server.Call(n + 1)
- parts := compactorInputParts(t, call)
- require.Equal(t, first, parts[:2])
- require.Equal(t, []string{compactorTooLong(want)}, parts[2:])
- require.Contains(t, parts[2], "was "+strconv.Itoa(len(want))+" bytes,")
- require.Contains(t, parts[2], "\n"+want[:512]+"| <- LIMIT")
-
- note, err := json.Marshal(map[string]string{
- "type": "input_text",
- "text": parts[2],
- })
+ 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")
+
+ note, err := json.Marshal(modelUserMessage(texts[2]))
require.NoError(t, err)
require.LessOrEqual(t, len(call.raw), len(server.Call(0).raw)+len(",")+len(note))
}
})
}
-// compactorInputParts returns the texts of the one user message of call.
-func compactorInputParts(t *testing.T, call ModelFakeCall) []string {
+// compactorInputTexts returns the text of each user message of call. Each
+// message has one text part.
+func compactorInputTexts(t *testing.T, call ModelFakeCall) []string {
t.Helper()
var body struct {
}
require.NoError(t, json.Unmarshal(call.raw, &body))
- require.Len(t, body.Input, 1)
- require.Equal(t, "user", body.Input[0].Role)
- texts := make([]string, len(body.Input[0].Content))
- for n, part := range body.Input[0].Content {
- require.Equal(t, "input_text", part.Type)
- texts[n] = part.Text
+ texts := make([]string, len(body.Input))
+ for n, message := range body.Input {
+ require.Equal(t, "user", message.Role)
+ require.Len(t, message.Content, 1)
+ require.Equal(t, "input_text", message.Content[0].Type)
+ texts[n] = message.Content[0].Text
}
return texts
projects / onechat / commitdiff