@pentatonic-ai/ai-agent-sdk 0.7.8 → 0.7.10

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pentatonic-ai/ai-agent-sdk",
-  "version": "0.7.8",
+  "version": "0.7.10",
   "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/src/ingest.js

@@ -41,6 +41,18 @@ import { distill } from "./distill.js";
  *   ids) — pass the raw form here so retries of the same logical event
  *   match across runs whose prefixes differ by a few ms. Defaults to
  *   `content`.
+ * @param {"client" | "user"} [opts.dedupScope="client"] - Scope of the
+ *   dedup match. Default `"client"`: byte-equal content for the tenant
+ *   collapses to one row regardless of who emitted it (today's behaviour,
+ *   appropriate when each row is single-owner). Set `"user"` to also
+ *   require `user_id` equality, which lets multiple users legitimately
+ *   own their own copy of the same shared content (private chat-channel
+ *   members, group meeting attendees) — emit one ingest per member with
+ *   the same content but different `userId`, all with
+ *   `dedupScope: "user"`, and each user gets their own row with native
+ *   per-user access counting / recency / decay. Requires `userId` to be
+ *   set; when `userId` is null the option degrades to `"client"` scope
+ *   (a global shared row still collapses cross-emit).
  * @param {number} [opts.dedupLegacyWindowDays=7] - How far back the
  *   `[<iso>] <content>` legacy-form `LIKE` match scans. Default 7 days.
  *   The leading-wildcard `LIKE` can't use a btree index, so without a
@@ -94,6 +106,11 @@ export async function ingest(db, ai, llm, content, opts = {}) {
       opts.dedupLegacyWindowDays === undefined
         ? 7
         : Number(opts.dedupLegacyWindowDays);
+    // Per-user dedup requires `userId`; degrade to client-scope otherwise so
+    // a misuse can't accidentally relax the dedup boundary (we'd rather over-
+    // collapse than fragment the corpus on a missing userId).
+    const userScopedDedup = opts.dedupScope === "user" && !!opts.userId;
+    const userClause = userScopedDedup ? ` AND user_id = $4` : "";
     try {
       const sql =
         legacyWindowDays > 0
@@ -107,17 +124,18 @@ export async function ingest(db, ai, llm, content, opts = {}) {
                     content LIKE '%] ' || $2
                     AND created_at > NOW() - ($3 || ' days')::interval
                   )
-                )
+                )${userScopedDedup ? "\n                AND user_id = $4" : ""}
               LIMIT 1`
           : `SELECT id, 'exact' AS match_kind
                FROM memory_nodes
               WHERE client_id = $1
-                AND content = $2
+                AND content = $2${userClause}
               LIMIT 1`;
-      const params =
+      const baseParams =
         legacyWindowDays > 0
           ? [clientId, dedupKey, String(legacyWindowDays)]
           : [clientId, dedupKey];
+      const params = userScopedDedup ? [...baseParams, opts.userId] : baseParams;
       const dupCheck = await db(sql, params);
       if (dupCheck.rows?.length) {
         const matchKind = dupCheck.rows[0].match_kind || "exact";

package/packages/memory-engine/compat/server.py

@@ -30,7 +30,6 @@ Environment:
     L6_DOC_URL               default http://l6:8037
     NV_EMBED_URL             default http://nv-embed:8041/v1/embeddings
     PORT                     default 8099 (matches pentatonic-memory v0.5)
-    CLIENT_ID                default "default"
 """
 
 import hashlib
@@ -63,7 +62,18 @@ NEO4J_AUTH = os.environ.get("NEO4J_AUTH", "neo4j/local-dev-pw")
 NEO4J_DB = os.environ.get("NEO4J_DB", "neo4j")
 
 PORT = int(os.environ.get("PORT", "8099"))
-CLIENT_ID = os.environ.get("CLIENT_ID", "default")
+
+
+# Layer types we surface as the SDK 4-layer projection. Engine stores
+# everything as chunks tagged with arena + layer_type metadata; this
+# helper renders the legacy `ml_<arena>_<type>` layer-id from the
+# per-row arena, so the response reflects the actual data not a
+# deployment-wide constant. Falls back to "episodic" when arena or
+# layer_type is missing.
+def _layer_id(arena: Optional[str], layer_type: Optional[str] = None) -> str:
+    a = arena or "general"
+    t = layer_type or "episodic"
+    return f"ml_{a}_{t}"
 
 # Test/isolated mode: bypass the L2 HybridRAG orchestrator and query L6 directly.
 # Useful for bench harnesses where you want to validate the ingest+search
@@ -384,7 +394,6 @@ async def health():
     """
     out = {
         "status": "ok",
-        "client": CLIENT_ID,
         "version": VERSION,
         "engine": "pentatonic-memory-engine",
         "layers": {},
@@ -476,7 +485,7 @@ async def store(req: StoreRequest):
     return {
         "id": rid,
         "content": req.content,
-        "layerId": f"ml_{CLIENT_ID}_episodic",
+        "layerId": _layer_id(arena, (req.metadata or {}).get("layer_type")),
         "engine": {
             "l0": l2_internal.get("l0", 0),
             "l3_chunks": l2_internal.get("l3_chunks", 0),
@@ -717,13 +726,14 @@ async def search(req: SearchRequest):
                 if item.get(k)
             }
             merged_meta = {**raw_top_level, **(attached_meta or item.get("metadata") or {})}
+            row_arena = merged_meta.get("arena")
+            row_layer_type = merged_meta.get("layer_type")
             out_results.append({
                 "id": key,
                 "content": item.get("text") or item.get("content") or item.get("snippet") or "",
                 "metadata": merged_meta,
                 "similarity": float(rrf_scores[key]),
-                "layer_id": f"ml_{CLIENT_ID}_episodic",
-                "client_id": CLIENT_ID,
+                "layer_id": _layer_id(row_arena, row_layer_type),
                 "source": item.get("source_file") or item.get("path") or "",
                 "engine_layer": "+".join(sorted(set(layer_provenance.get(key, [])))),
             })
@@ -819,13 +829,14 @@ async def search(req: SearchRequest):
             if item.get(k)
         }
         merged_meta = {**raw_top_level, **(attached_meta or item.get("metadata") or {})}
+        row_arena = merged_meta.get("arena")
+        row_layer_type = merged_meta.get("layer_type")
         out_results.append({
             "id": chosen_id,
             "content": item.get("text") or item.get("content") or item.get("snippet") or "",
             "metadata": merged_meta,
             "similarity": float(item.get("score") or item.get("similarity") or 0.0),
-            "layer_id": f"ml_{CLIENT_ID}_episodic",
-            "client_id": CLIENT_ID,
+            "layer_id": _layer_id(row_arena, row_layer_type),
             "source": item.get("source", item.get("source_file", "")),
             "engine_layer": item.get("layer", item.get("source_layer", "")),
         })

package/packages/memory-engine/docker-compose.yml

@@ -215,7 +215,6 @@ services:
       L5_MILVUS_URL: http://l5:8034
       L6_DOC_URL: http://l6:8037
       NV_EMBED_URL: ${NV_EMBED_URL:-http://host.docker.internal:8041/v1/embeddings}
-      CLIENT_ID: ${CLIENT_ID:-default}
       BYPASS_L2_PROXY: ${BYPASS_L2_PROXY:-0}
     extra_hosts:
       - "host.docker.internal:host-gateway"