crispy-recall 0.1.2 → 0.1.4

package/dist/embed-pending.js

@@ -57,6 +57,9 @@ function runDir() {
 function ensureDir() {
   (0, import_node_fs.mkdirSync)(recallRoot(), { recursive: true });
 }
+function transcriptGlob(...segments) {
+  return segments.join("/").replace(/\\/g, "/");
+}
 var import_node_fs, import_node_os, import_node_path, rootOverride;
 var init_paths = __esm({
   "src/paths.ts"() {
@@ -4594,6 +4597,11 @@ async function startServer() {
     "-ub",
     "8192",
     // micro-batch (ubatch) — also defaults to 512, must be raised
+    // Flash Attention: the attention compute buffer is otherwise O(n²) in the
+    // 8192-token batch (~3.7 GB on GPU), which OOMs cards <6 GB and forces CPU.
+    // -fa makes it O(n) (~0.8 GB → ~2.9 GB total), so 4 GB cards run on GPU,
+    // with numerically identical embeddings (verified cosine 1.0 vs non-FA).
+    "-fa",
     // nomic-embed-text-v1.5 is trained at 2048 ctx and uses Dynamic NTK-aware
     // RoPE scaling to extend to 8192. Without these flags, newer llama.cpp
     // (b9253+) refuses inputs >2048 tokens, and older versions silently
@@ -4801,6 +4809,9 @@ async function embedViaProcess(texts, modelPath) {
       "array",
       "-c",
       "8192",
+      // Flash Attention — see startServer: O(n) attention buffer instead of
+      // O(n²), so the one-shot path also fits cards <6 GB. Identical embeddings.
+      "-fa",
       // Match the YaRN flags on the llama-server batch path — nomic-embed-text-v1.5
       // is trained at 2048 ctx and extends to 8192 via Dynamic NTK-aware RoPE.
       // Without these, newer llama.cpp (b9253+) refuses inputs >2048 tokens here
@@ -13195,9 +13206,9 @@ async function mtimeScan(opts) {
   const codexRoot = process.env["CODEX_HOME"] ?? (0, import_node_path7.join)((0, import_node_os3.homedir)(), ".codex");
   const patterns = [];
   if (vendors.includes("claude"))
-    patterns.push([(0, import_node_path7.join)(claudeRoot, "projects", "**", "*.jsonl"), "claude"]);
+    patterns.push([transcriptGlob(claudeRoot, "projects", "**", "*.jsonl"), "claude"]);
   if (vendors.includes("codex"))
-    patterns.push([(0, import_node_path7.join)(codexRoot, "sessions", "**", "*.jsonl"), "codex"]);
+    patterns.push([transcriptGlob(codexRoot, "sessions", "**", "*.jsonl"), "codex"]);
   const db3 = getDb(dbPath());
   const watermarks = /* @__PURE__ */ new Map();
   const rows = db3.all("SELECT * FROM ingest_watermark");

package/dist/recall.js

@@ -2570,6 +2570,9 @@ function logsDir() {
 function ensureDir() {
   (0, import_node_fs.mkdirSync)(recallRoot(), { recursive: true });
 }
+function transcriptGlob(...segments) {
+  return segments.join("/").replace(/\\/g, "/");
+}
 var import_node_fs, import_node_os, import_node_path, rootOverride;
 var init_paths = __esm({
   "src/paths.ts"() {
@@ -3268,6 +3271,11 @@ async function startServer() {
     "-ub",
     "8192",
     // micro-batch (ubatch) — also defaults to 512, must be raised
+    // Flash Attention: the attention compute buffer is otherwise O(n²) in the
+    // 8192-token batch (~3.7 GB on GPU), which OOMs cards <6 GB and forces CPU.
+    // -fa makes it O(n) (~0.8 GB → ~2.9 GB total), so 4 GB cards run on GPU,
+    // with numerically identical embeddings (verified cosine 1.0 vs non-FA).
+    "-fa",
     // nomic-embed-text-v1.5 is trained at 2048 ctx and uses Dynamic NTK-aware
     // RoPE scaling to extend to 8192. Without these flags, newer llama.cpp
     // (b9253+) refuses inputs >2048 tokens, and older versions silently
@@ -3475,6 +3483,9 @@ async function embedViaProcess(texts, modelPath) {
       "array",
       "-c",
       "8192",
+      // Flash Attention — see startServer: O(n) attention buffer instead of
+      // O(n²), so the one-shot path also fits cards <6 GB. Identical embeddings.
+      "-fa",
       // Match the YaRN flags on the llama-server batch path — nomic-embed-text-v1.5
       // is trained at 2048 ctx and extends to 8192 via Dynamic NTK-aware RoPE.
       // Without these, newer llama.cpp (b9253+) refuses inputs >2048 tokens here
@@ -13749,9 +13760,9 @@ async function mtimeScan(opts) {
   const codexRoot = process.env["CODEX_HOME"] ?? (0, import_node_path6.join)((0, import_node_os3.homedir)(), ".codex");
   const patterns = [];
   if (vendors.includes("claude"))
-    patterns.push([(0, import_node_path6.join)(claudeRoot, "projects", "**", "*.jsonl"), "claude"]);
+    patterns.push([transcriptGlob(claudeRoot, "projects", "**", "*.jsonl"), "claude"]);
   if (vendors.includes("codex"))
-    patterns.push([(0, import_node_path6.join)(codexRoot, "sessions", "**", "*.jsonl"), "codex"]);
+    patterns.push([transcriptGlob(codexRoot, "sessions", "**", "*.jsonl"), "codex"]);
   const db4 = getDb(dbPath());
   const watermarks = /* @__PURE__ */ new Map();
   const rows = db4.all("SELECT * FROM ingest_watermark");
@@ -14548,7 +14559,7 @@ function listAllSessions(opts) {
   const codexRoot = process.env["CODEX_HOME"] ?? (0, import_node_path7.join)((0, import_node_os4.homedir)(), ".codex");
   const out = [];
   if (vendors.includes("claude")) {
-    const files = globSync((0, import_node_path7.join)(claudeRoot, "projects", "**", "*.jsonl"), { nodir: true });
+    const files = globSync(transcriptGlob(claudeRoot, "projects", "**", "*.jsonl"), { nodir: true });
     for (const file of files) {
       out.push({
         sessionId: sessionIdFromPath2(file, "claude"),
@@ -14559,7 +14570,7 @@ function listAllSessions(opts) {
     }
   }
   if (vendors.includes("codex")) {
-    const files = globSync((0, import_node_path7.join)(codexRoot, "sessions", "**", "*.jsonl"), { nodir: true });
+    const files = globSync(transcriptGlob(codexRoot, "sessions", "**", "*.jsonl"), { nodir: true });
     for (const file of files) {
       out.push({
         sessionId: sessionIdFromPath2(file, "codex"),
@@ -14585,6 +14596,7 @@ var init_session_manager_shim = __esm({
     init_esm5();
     import_node_os4 = require("node:os");
     import_node_path7 = require("node:path");
+    init_paths();
   }
 });
 
@@ -15016,6 +15028,12 @@ async function defaultProbe(args) {
     "array",
     "-c",
     "8192",
+    // Probe with the SAME flags the real embed server uses (esp. -fa). Without
+    // Flash Attention the probe allocates the ~3.7 GB O(n²) attention buffer and
+    // OOMs on a 4 GB card — failing the probe and falling back to CPU even
+    // though the actual -fa server (~2.9 GB) would have fit. Probe must mirror
+    // runtime to gate GPU adoption correctly.
+    "-fa",
     "--rope-scaling",
     "yarn",
     "--rope-freq-scale",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "crispy-recall",
-  "version": "0.1.2",
+  "version": "0.1.4",
   "description": "Local session transcript memory for Claude Code and Codex — search past sessions with FTS5 + semantic vectors.",
   "license": "MIT",
   "author": "Sylvester Wong",