"encoding/json"
"errors"
"fmt"
+ "log/slog"
"os"
"path/filepath"
"strings"
errMemoryUnsettled = errors.New("a summary of the view failed, and it is tried again at the next message")
)
+// Summarizer builds the text of one node with a model.
+type Summarizer interface {
+ Summarize(ctx context.Context, task Task) (string, error)
+}
+
+// Memory is the memory of one chat: the message log, the summary tree, the chat
+// view, the compaction view, and the queue of the builds. The compaction view
+// is the chat view merged further, at the limits of 16 to 32 KB. It takes the
+// same new lines, and it merges again in one batch when it is too large or when
+// the chat view merges. So the compactions share its prefix in the prompt
+// cache. It saves the views to view.json after each change and loads them at
+// start. It never builds the views again from the log, because a new view
+// differs from the live one and so loses the prompt cache.
+//
+// Each write goes to the operating system at once, and a syncer makes the
+// writes durable by group commit in the background. Flush waits until the
+// writes so far are durable.
+//
+// The summarizer builds the nodes in the queue in the background. The start
+// reads the indexes once to fill the queue again. After that, only a new
+// message or a new node adds work. Memory keeps the nodes of the views and the
+// tasks of the running builds. It reads all other messages and nodes from their
+// day files through the indexes.
+type Memory struct {
+ mutex sync.Mutex
+ root string
+ log *Log
+ syncer *Syncer
+ tree *Tree
+ chat *View
+ cview *View
+ queue *BuildQueue
+ summarizer Summarizer
+ logger *slog.Logger
+ ctx context.Context
+ cancel context.CancelFunc
+ workers sync.WaitGroup
+ changed chan struct{}
+}
+
+// Task is the build of the node (Level, Index). Context is the compaction view
+// for the build. Input holds the source "kind: text" of a leaf, or the two
+// child lines of a merge with their newlines as spaces. Images holds the leaf
+// image filenames and is empty for merges.
+type Task struct {
+ Level int
+ Index int
+ Context string
+ Input []string
+ Images []string
+}
+
+// MemoryViewFile is the content of view.json.
+type MemoryViewFile struct {
+ Chat ViewState `json:"chat"`
+ Compaction ViewState `json:"compaction"`
+}
+
// openMemory opens the chat in root and starts the builds that are ready. The
-// clock gives the dates of new messages and the days of new nodes. The builds
-// stop when ctx ends or at Close.
-func openMemory(ctx context.Context, root string, clock func() time.Time, summarizer Summarizer) (*Memory, error) {
+// clock gives the dates of new messages and the days of new nodes. logger gets
+// one line for each failed build. The builds stop when ctx ends or at Close.
+func openMemory(ctx context.Context, root string, clock func() time.Time, summarizer Summarizer, logger *slog.Logger) (*Memory, error) {
log, err := openLog(ctx, root, clock)
if err != nil {
return nil, err
log: log,
queue: newBuildQueue(),
summarizer: summarizer,
+ logger: logger,
changed: make(chan struct{}),
}
return m, nil
}
-// Summarizer builds the text of one node with a model.
-type Summarizer interface {
- Summarize(ctx context.Context, task Task) (string, error)
-}
-
-// Memory is the memory of one chat: the message log, the summary tree, the chat
-// view, the compaction view, and the queue of the builds. The compaction view
-// is the chat view merged further, at the limits of 16 to 32 KB. It takes the
-// same new lines, and it merges again in one batch when it is too large or when
-// the chat view merges. So the compactions share its prefix in the prompt
-// cache. It saves the views to view.json after each change and loads them at
-// start. It never builds the views again from the log, because a new view
-// differs from the live one and so loses the prompt cache.
-//
-// Each write goes to the operating system at once, and a syncer makes the
-// writes durable by group commit in the background. Flush waits until the
-// writes so far are durable.
-//
-// The summarizer builds the nodes in the queue in the background. The start
-// reads the indexes once to fill the queue again. After that, only a new
-// message or a new node adds work. Memory keeps the nodes of the views and the
-// tasks of the running builds. It reads all other messages and nodes from their
-// day files through the indexes.
-type Memory struct {
- mutex sync.Mutex
- root string
- log *Log
- syncer *Syncer
- tree *Tree
- chat *View
- cview *View
- queue *BuildQueue
- summarizer Summarizer
- ctx context.Context
- cancel context.CancelFunc
- workers sync.WaitGroup
- changed chan struct{}
-}
-
// Close stops the builds, waits for them, makes the writes durable, closes the
// files of the chat, and releases its lock.
func (m *Memory) Close() error {
defer m.mutex.Unlock()
var err error
+
if m.syncer != nil {
err = m.syncer.Close()
}
// Append logs a new message, adds it to the tree and the views, and starts the
// builds that are ready, also the failed ones.
func (m *Memory) Append(kind, text string) (Message, error) {
+ return m.AppendDated(kind, text, "")
+}
+
+// AppendDated appends a message with its RFC 3339 date and builds its nodes. An
+// empty date uses the clock, as Append does.
+func (m *Memory) AppendDated(kind, text, date string) (Message, error) {
m.mutex.Lock()
defer m.mutex.Unlock()
- message, err := m.log.Append(kind, text)
+ return m.appendDated(kind, text, date)
+}
+
+// appendDated appends under the memory lock and updates the views and builds.
+func (m *Memory) appendDated(kind, text, date string) (Message, error) {
+ message, err := m.log.AppendDated(kind, text, date)
if err != nil {
- return Message{}, err
+ return Message{}, fmt.Errorf("append dated log message: %w", err)
}
defer m.syncer.Written()
err = m.addMessage(message)
if err != nil {
- return Message{}, err
+ return Message{}, fmt.Errorf("add dated message to memory: %w", err)
}
err = m.save()
if err != nil {
- return Message{}, err
+ return Message{}, fmt.Errorf("save dated message view: %w", err)
}
m.queue.Retry()
return message, nil
}
+// AppendOnce appends a queued message. prepare saves its log ID before the
+// append, under the memory lock. A saved ID permits recovery after a stop
+// between that save and the append, without a second log entry. prepare must
+// not call Memory. The caller stops delivery after a failed append.
+func (m *Memory) AppendOnce(kind, text string, id *int, prepare func(int) error) (Message, error) {
+ m.mutex.Lock()
+ defer m.mutex.Unlock()
+
+ if id != nil && *id < m.log.Len() {
+ message, err := m.log.Message(*id)
+ if err != nil {
+ return Message{}, fmt.Errorf("read queued message: %w", err)
+ }
+
+ if message.Kind != kind || message.Text != text {
+ return Message{}, errors.New("queued message does not match its log entry")
+ }
+
+ return message, nil
+ }
+
+ if id != nil && *id != m.log.Len() {
+ return Message{}, errors.New("queued message is beyond the log")
+ }
+
+ err := prepare(m.log.Len())
+ if err != nil {
+ return Message{}, fmt.Errorf("prepare queued message: %w", err)
+ }
+
+ message, err := m.appendDated(kind, text, "")
+ if err != nil {
+ return Message{}, fmt.Errorf("append queued message: %w", err)
+ }
+
+ return message, nil
+}
+
// children reads the two children of the node key.
func (m *Memory) children(key NodeKey) (Node, Node, error) {
var children [2]Node
+
for n := range children {
child, built, err := m.tree.Node(key.level-1, 2*key.index+n)
if err != nil {
func (m *Memory) Wait(ctx context.Context) error {
for {
m.mutex.Lock()
+
running := m.queue.Running()
changed := m.changed
m.mutex.Unlock()
func (m *Memory) Settle(ctx context.Context) error {
for {
m.mutex.Lock()
+
pending := m.chat.Pending()
running := m.queue.Running()
changed := m.changed
return m.syncer.Flush(ctx)
}
+// Snapshot flushes all memory writes and runs snapshot while no append or
+// summary write can change the log, indexes, tree, or views. snapshot must not
+// call Memory.
+func (m *Memory) Snapshot(ctx context.Context, snapshot func() error) error {
+ m.mutex.Lock()
+ defer m.mutex.Unlock()
+
+ err := m.Flush(ctx)
+ if err != nil {
+ return fmt.Errorf("flush memory snapshot: %w", err)
+ }
+
+ err = snapshot()
+ if err != nil {
+ return fmt.Errorf("copy memory snapshot: %w", err)
+ }
+
+ return nil
+}
+
// ViewLines returns the lines of the chat view, each with its newline, without
// the leaf lines of the messages in skip.
func (m *Memory) ViewLines(skip map[int]bool) []string {
}
var file MemoryViewFile
+
err = json.Unmarshal(data, &file)
if err != nil {
return fmt.Errorf("decode view: %w", err)
leaf, free := freeLeaf(message)
if !free {
m.queue.AddLeaf(id)
+
continue
}
task, err := m.task(key)
if err != nil {
+ m.failed(key, fmt.Errorf("read its input: %w", err))
m.queue.Fail(key)
+
continue
}
m.workers.Add(1)
+
go m.run(key, task)
}
}
// addMessage builds the leaf of message when it needs no model call, or else
-// adds the leaf to the queue. Then it appends the leaf to the views. A crash can
-// stop an append after the leaf build and before the view save, so a leaf that
-// is built already stays as it is, and restoreQueue settles its parent.
+// adds the leaf to the queue. Then it appends the leaf to the views. A crash
+// can stop an append after the leaf build and before the view save, so a leaf
+// that is built already stays as it is, and restoreQueue settles its parent.
func (m *Memory) addMessage(message Message) error {
built, err := m.tree.Built(0, message.ID)
if err != nil {
node, free := freeMerge(left, right)
if !free {
m.queue.AddMerge(parent)
+
return nil
}
}
task.Input = []string{leafSource(message)}
+ task.Images = imageFiles(message.Text)
return task, nil
}
Text: text,
Size: len(text),
})
+
m.syncer.Written()
}
+ if err != nil && m.ctx.Err() == nil {
+ m.failed(key, err)
+ }
+
if err != nil {
m.queue.Fail(key)
} else {
}
close(m.changed)
+
m.changed = make(chan struct{})
m.dispatch()
}
+// failed logs the failed build of the node key with its cause err.
+func (m *Memory) failed(key NodeKey, err error) {
+ m.logger.Warn("summary failed", "level", key.level, "index", key.index, "err", err)
+}
+
// ready tells if both children of the node key are built and the node is not.
func (m *Memory) ready(key NodeKey) (bool, error) {
for _, index := range []int{2 * key.index, 2*key.index + 1} {
return m.context(level, index)
}
-// Task is the build of the node (Level, Index). Context is the compaction view
-// for the build. Input holds the source "kind: text" of a leaf, or the two
-// child lines of a merge with their newlines as spaces.
-type Task struct {
- Level int
- Index int
- Context string
- Input []string
+// rebuildComplete checks the whole tree, not only the visible view. A compactor
+// can fail without an HTTP error, for example with an empty line. Such a failed
+// merge can be absent from the view and must prevent a swap.
+func (m *Memory) rebuildComplete() error {
+ m.mutex.Lock()
+ defer m.mutex.Unlock()
+
+ if len(m.queue.leaves) != 0 || len(m.queue.queued) != 0 {
+ return errors.New("rebuild failed: summary tree has unfinished nodes")
+ }
+
+ return nil
}
-// MemoryViewFile is the content of view.json.
-type MemoryViewFile struct {
- Chat ViewState `json:"chat"`
- Compaction ViewState `json:"compaction"`
+// buildModelLines returns the sizes of the nodes that the model wrote, in tree
+// order. A node that fits its source with no model call is not one of them.
+func (m *Memory) buildModelLines() ([]int, error) {
+ m.mutex.Lock()
+ defer m.mutex.Unlock()
+
+ keys, err := m.tree.Keys()
+ if err != nil {
+ return nil, fmt.Errorf("list built nodes: %w", err)
+ }
+
+ var sizes []int
+
+ for _, key := range keys {
+ var node Node
+
+ node, _, err = m.tree.Node(key.level, key.index)
+ if err != nil {
+ return nil, fmt.Errorf("read built node: %w", err)
+ }
+
+ var model bool
+
+ model, err = m.buildBuiltByModel(key)
+ if err != nil {
+ return nil, fmt.Errorf("check model source: %w", err)
+ }
+
+ if model {
+ sizes = append(sizes, node.Size)
+ }
+ }
+
+ return sizes, nil
+}
+
+// buildBuiltByModel tells if the node key needed a model call: a leaf whose
+// source is over nodeBytes, or a merge whose children do not fit in one node.
+func (m *Memory) buildBuiltByModel(key NodeKey) (bool, error) {
+ if key.level == 0 {
+ message, err := m.log.Message(key.index)
+ if err != nil {
+ return false, fmt.Errorf("read leaf source: %w", err)
+ }
+
+ _, free := freeLeaf(message)
+
+ return !free, nil
+ }
+
+ left, right, err := m.children(key)
+ if err != nil {
+ return false, fmt.Errorf("read merge sources: %w", err)
+ }
+
+ _, free := freeMerge(left, right)
+
+ return !free, nil
}
projects / onechat / blobdiff