p2: digest with in-sift FIFO, SQLite caches, Honcho context, recover

Dixie is capped (max_inflight 2, queue_depth 8, queue_wait 20s) and
never queued on llama-server. Streaming chat, thinking off, 45s
fallback to clean. Digest and chunk-embedding caches are
content-addressed SQLite (Honcho context is not cached). Recover
uses the guarded Fetcher. guardian-star is refused in code.
This commit is contained in:
2026-09-22 23:54:49 -07:00
parent 505e113b64
commit 2f9d3d97e6
2023 changed files with 3815943 additions and 30 deletions
+229
View File
@@ -0,0 +1,229 @@
// Package dcache is the content-addressed digest + chunk-embedding SQLite cache.
// Honcho context is never stored here.
package dcache
import (
"crypto/sha256"
"database/sql"
"encoding/binary"
"encoding/hex"
"fmt"
"log/slog"
"math"
"os"
"strings"
"sync"
"time"
_ "modernc.org/sqlite"
)
const PromptVersion = "1"
type DB struct {
sql *sql.DB
maxBytes int64
log *slog.Logger
mu sync.Mutex
DigestH uint64
DigestM uint64
EmbedH uint64
EmbedM uint64
Evictions uint64
}
func Open(path string, maxBytes int64, log *slog.Logger) (*DB, error) {
if maxBytes <= 0 {
maxBytes = 500 << 20
}
if err := os.MkdirAll(dirOf(path), 0o750); err != nil && path != ":memory:" && path != "" {
// ignore: sqlite will fail if it cannot create
}
dsn := path
if path == "" {
dsn = ":memory:"
}
sq, err := sql.Open("sqlite", dsn)
if err != nil {
return nil, err
}
sq.SetMaxOpenConns(1)
if _, err := sq.Exec(`
PRAGMA journal_mode=WAL;
CREATE TABLE IF NOT EXISTS digest (
k TEXT PRIMARY KEY,
text TEXT NOT NULL,
chunks_used INTEGER,
last_hit INTEGER NOT NULL,
nbytes INTEGER NOT NULL
);
CREATE TABLE IF NOT EXISTS embed (
k TEXT PRIMARY KEY,
vec BLOB NOT NULL,
last_hit INTEGER NOT NULL,
nbytes INTEGER NOT NULL
);
`); err != nil {
_ = sq.Close()
return nil, err
}
d := &DB{sql: sq, maxBytes: maxBytes, log: log}
return d, nil
}
func (d *DB) Close() error {
if d == nil || d.sql == nil {
return nil
}
return d.sql.Close()
}
func NormalizeQuestion(q string) string {
return strings.ToLower(strings.Join(strings.Fields(q), " "))
}
func DigestKey(contentHash, question, model, promptVer string) string {
sum := sha256.Sum256([]byte(contentHash + "\x00" + NormalizeQuestion(question) + "\x00" + model + "\x00" + promptVer))
return hex.EncodeToString(sum[:])
}
func EmbedKey(contentHash string, chunkIndex int, embedModel string) string {
sum := sha256.Sum256([]byte(fmt.Sprintf("%s\x00%d\x00%s", contentHash, chunkIndex, embedModel)))
return hex.EncodeToString(sum[:])
}
func QuestionEmbedKey(question, embedModel string) string {
sum := sha256.Sum256([]byte("q\x00" + NormalizeQuestion(question) + "\x00" + embedModel))
return hex.EncodeToString(sum[:])
}
type DigestHit struct {
Text string
ChunksUsed int
}
func (d *DB) GetDigest(key string) (DigestHit, bool) {
if d == nil {
return DigestHit{}, false
}
var h DigestHit
var n int
err := d.sql.QueryRow(`SELECT text, chunks_used FROM digest WHERE k = ?`, key).Scan(&h.Text, &n)
if err != nil {
d.mu.Lock()
d.DigestM++
d.mu.Unlock()
return DigestHit{}, false
}
h.ChunksUsed = n
_, _ = d.sql.Exec(`UPDATE digest SET last_hit = ? WHERE k = ?`, time.Now().Unix(), key)
d.mu.Lock()
d.DigestH++
d.mu.Unlock()
return h, true
}
func (d *DB) PutDigest(key, text string, chunksUsed int) {
if d == nil {
return
}
n := len(key) + len(text)
_, err := d.sql.Exec(`INSERT OR REPLACE INTO digest(k, text, chunks_used, last_hit, nbytes) VALUES(?,?,?,?,?)`,
key, text, chunksUsed, time.Now().Unix(), n)
if err != nil && d.log != nil {
d.log.Warn("digest cache write", "err", err)
}
d.evict()
}
func (d *DB) GetEmbed(key string) ([]float64, bool) {
if d == nil {
return nil, false
}
var b []byte
err := d.sql.QueryRow(`SELECT vec FROM embed WHERE k = ?`, key).Scan(&b)
if err != nil {
d.mu.Lock()
d.EmbedM++
d.mu.Unlock()
return nil, false
}
_, _ = d.sql.Exec(`UPDATE embed SET last_hit = ? WHERE k = ?`, time.Now().Unix(), key)
d.mu.Lock()
d.EmbedH++
d.mu.Unlock()
return blobToVec(b), true
}
func (d *DB) PutEmbed(key string, vec []float64) {
if d == nil {
return
}
b := vecToBlob(vec)
_, err := d.sql.Exec(`INSERT OR REPLACE INTO embed(k, vec, last_hit, nbytes) VALUES(?,?,?,?)`,
key, b, time.Now().Unix(), len(key)+len(b))
if err != nil && d.log != nil {
d.log.Warn("embed cache write", "err", err)
}
d.evict()
}
func (d *DB) Size() int64 {
if d == nil {
return 0
}
var n sql.NullInt64
_ = d.sql.QueryRow(`SELECT IFNULL(SUM(nbytes),0) FROM digest`).Scan(&n)
var m sql.NullInt64
_ = d.sql.QueryRow(`SELECT IFNULL(SUM(nbytes),0) FROM embed`).Scan(&m)
return n.Int64 + m.Int64
}
func (d *DB) evict() {
if d == nil {
return
}
for d.Size() > d.maxBytes {
var k string
err := d.sql.QueryRow(`
SELECT k FROM (
SELECT k, last_hit, 0 AS kind FROM digest
UNION ALL
SELECT k, last_hit, 1 FROM embed
) ORDER BY last_hit ASC LIMIT 1`).Scan(&k)
if err != nil {
return
}
_, _ = d.sql.Exec(`DELETE FROM digest WHERE k = ?`, k)
_, _ = d.sql.Exec(`DELETE FROM embed WHERE k = ?`, k)
d.mu.Lock()
d.Evictions++
d.mu.Unlock()
}
}
func vecToBlob(v []float64) []byte {
b := make([]byte, 4*len(v))
for i, x := range v {
binary.LittleEndian.PutUint32(b[i*4:], math.Float32bits(float32(x)))
}
return b
}
func blobToVec(b []byte) []float64 {
n := len(b) / 4
out := make([]float64, n)
for i := 0; i < n; i++ {
out[i] = float64(math.Float32frombits(binary.LittleEndian.Uint32(b[i*4:])))
}
return out
}
func dirOf(path string) string {
i := strings.LastIndex(path, "/")
if i <= 0 {
return "."
}
return path[:i]
}
+87
View File
@@ -0,0 +1,87 @@
package dcache
import (
"os"
"path/filepath"
"testing"
)
func TestHitMissAndNormalise(t *testing.T) {
d, err := Open(filepath.Join(t.TempDir(), "c.db"), 1<<20, nil)
if err != nil {
t.Fatal(err)
}
defer d.Close()
k1 := DigestKey("abc", "What Share?", "m", PromptVersion)
k2 := DigestKey("abc", "what share?", "m", PromptVersion)
if k1 != k2 {
t.Fatal("normalise")
}
if _, ok := d.GetDigest(k1); ok {
t.Fatal("miss")
}
d.PutDigest(k1, "forty two", 1)
h, ok := d.GetDigest(k2)
if !ok || h.Text != "forty two" {
t.Fatalf("%v %v", ok, h)
}
k3 := DigestKey("abd", "what share?", "m", PromptVersion)
if _, ok := d.GetDigest(k3); ok {
t.Fatal("changed content must miss")
}
}
func TestPromptVersionInvalidates(t *testing.T) {
k1 := DigestKey("h", "q", "m", "1")
k2 := DigestKey("h", "q", "m", "2")
if k1 == k2 {
t.Fatal("version")
}
}
func TestNoContextTable(t *testing.T) {
d, err := Open(filepath.Join(t.TempDir(), "c.db"), 1<<20, nil)
if err != nil {
t.Fatal(err)
}
defer d.Close()
var n int
_ = d.sql.QueryRow(`SELECT COUNT(*) FROM sqlite_master WHERE type='table' AND name='context'`).Scan(&n)
if n != 0 {
t.Fatal("must not store context")
}
}
func TestEvictUnderCap(t *testing.T) {
d, err := Open(filepath.Join(t.TempDir(), "c.db"), 200, nil)
if err != nil {
t.Fatal(err)
}
defer d.Close()
for i := 0; i < 20; i++ {
d.PutDigest(DigestKey("h", string(rune('a'+i)), "m", "1"), "xxxxxxxxxxxxxxxxxxxx", 1)
}
if d.Size() > 400 {
t.Fatalf("size %d", d.Size())
}
if d.Evictions == 0 {
t.Fatal("want evictions")
}
}
func TestCorruptDBBypass(t *testing.T) {
p := filepath.Join(t.TempDir(), "c.db")
if err := os.WriteFile(p, []byte("not a database"), 0o600); err != nil {
t.Fatal(err)
}
d, err := Open(p, 1<<20, nil)
if err == nil {
_ = d.Close()
// Open may succeed then query fails — GetDigest must miss not panic
if _, ok := d.GetDigest("x"); ok {
t.Fatal("corrupt hit")
}
return
}
// open failed: caller warns and bypasses
}