@pentatonic-ai/ai-agent-sdk 0.10.4 → 0.10.5

package/dist/index.cjs

@@ -878,7 +878,7 @@ function fireAndForgetEmit(clientConfig, sessionOpts, messages, result, model) {
 }
 
 // src/telemetry.js
-var VERSION = "0.10.4";
+var VERSION = "0.10.5";
 var TELEMETRY_URL = "https://sdk-telemetry.philip-134.workers.dev";
 function machineId() {
   const raw = typeof process !== "undefined" ? `${process.env?.USER || process.env?.USERNAME || "u"}:${process.platform || "x"}:${process.arch || "x"}` : "browser";

package/dist/index.js

@@ -847,7 +847,7 @@ function fireAndForgetEmit(clientConfig, sessionOpts, messages, result, model) {
 }
 
 // src/telemetry.js
-var VERSION = "0.10.4";
+var VERSION = "0.10.5";
 var TELEMETRY_URL = "https://sdk-telemetry.philip-134.workers.dev";
 function machineId() {
   const raw = typeof process !== "undefined" ? `${process.env?.USER || process.env?.USERNAME || "u"}:${process.platform || "x"}:${process.arch || "x"}` : "browser";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pentatonic-ai/ai-agent-sdk",
-  "version": "0.10.4",
+  "version": "0.10.5",
   "description": "TES SDK — LLM observability and lifecycle tracking via Pentatonic Thing Event System. Track token usage, tool calls, and conversations. Manage things through event-sourced lifecycle stages with AI enrichment and vector search.",
   "type": "module",
   "main": "./dist/index.cjs",

package/packages/memory-engine-v2/extractor-async/worker.py

@@ -761,6 +761,25 @@ SKIP_ATTRIBUTE_SOURCES = set(
 )
 DISTILL_MAX_AGE_DAYS = int(os.environ.get("DISTILL_MAX_AGE_DAYS", "90"))
 
+# Layer-1 content pre-filter (cascade tier 1 — the cheap deterministic gate in
+# front of the student/7B). Skips events with NO extractable signal BEFORE the
+# LLM, so GPU is spent only on text that can yield facts.
+#   - bytes-garbage: a binary doc (raw PDF bytes) stored as text decodes to a
+#     wall of U+FFFD replacement chars. `build_event_block` feeds `content` to
+#     the LLM, so it sees the garbage and extracts nothing (live 2026-06-10:
+#     35,296/39,453 pentatonic-team `doc` events are bytes-garbage). HIGH
+#     PRECISION — real text effectively never crosses a 5–10% replacement-char
+#     ratio, so this is a zero-quality-loss skip. (Durable fix = extract PDF
+#     text at INGEST; this stops the GPU waste meanwhile.)
+#   - too-short: trivially short content (one-line acks / emoji) has no facts.
+#     Conservative and OFF by default (0) to guarantee zero quality loss; tune
+#     up once layer-2 (the student model) owns the borderline cases.
+SKIP_BYTES_GARBAGE = os.environ.get(
+    "DISTILL_SKIP_BYTES_GARBAGE", "true"
+).strip().lower() not in ("false", "0", "no", "off")
+GARBAGE_CHAR_RATIO = float(os.environ.get("DISTILL_GARBAGE_CHAR_RATIO", "0.10"))
+MIN_CONTENT_CHARS = int(os.environ.get("DISTILL_MIN_CONTENT_CHARS", "0"))
+
 
 def claim_next_batch(conn: psycopg.Connection) -> list[dict[str, Any]]:
     """Atomically claim up to BATCH_SIZE pending items. SKIP LOCKED so
@@ -843,6 +862,43 @@ def claim_next_batch(conn: psycopg.Connection) -> list[dict[str, Any]]:
             """,
             (DISTILL_MAX_AGE_DAYS, DISTILL_MAX_AGE_DAYS),
         )
+        # Pre-filter: bytes-garbage content. A binary doc (raw PDF bytes)
+        # stored as text decodes to mostly U+FFFD (chr(65533)); the LLM
+        # extracts nothing from it. Skip when the replacement-char ratio
+        # exceeds GARBAGE_CHAR_RATIO — real text never crosses it, so no
+        # quality loss. Scoped to the pending set; one cheap UPDATE/cycle.
+        if SKIP_BYTES_GARBAGE:
+            cur.execute(
+                """
+                UPDATE distillation_queue dq SET
+                  status = 'done',
+                  completed_at = NOW(),
+                  last_error = 'filtered: bytes_garbage'
+                FROM events e
+                WHERE dq.event_id = e.id
+                  AND dq.status = 'pending'
+                  AND length(e.content) > 0
+                  AND (length(e.content) - length(replace(e.content, chr(65533), '')))::float
+                      / length(e.content) > %s
+                """,
+                (GARBAGE_CHAR_RATIO,),
+            )
+        # Pre-filter: trivially-short content (one-line acks / emoji). OFF by
+        # default (MIN_CONTENT_CHARS=0) so it never costs a fact unless tuned on.
+        if MIN_CONTENT_CHARS > 0:
+            cur.execute(
+                """
+                UPDATE distillation_queue dq SET
+                  status = 'done',
+                  completed_at = NOW(),
+                  last_error = 'filtered: too_short'
+                FROM events e
+                WHERE dq.event_id = e.id
+                  AND dq.status = 'pending'
+                  AND length(trim(e.content)) < %s
+                """,
+                (MIN_CONTENT_CHARS,),
+            )
 
     with conn.cursor(row_factory=psycopg.rows.dict_row) as cur:
         cur.execute(