@pentatonic-ai/ai-agent-sdk 0.8.3 → 0.8.4

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pentatonic-ai/ai-agent-sdk",
-  "version": "0.8.3",
+  "version": "0.8.4",
   "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/compat/server.py

@@ -453,21 +453,95 @@ async def health():
     if failures:
         out["status"] = "degraded" if failures < 3 else "down"
 
-    # Memory count: query L6 doc-store as authoritative. L6 /stats
-    # returns vector_chunks (Milvus) and fts_chunks (sqlite content
-    # table). Under healthy operation they're equal — take the max so
-    # the count is honest if one side is mid-rebuild.
+    # Per-layer chunk counts. Replaces the previous single `memories` int
+    # which only reflected L6's vector chunk count — misleading because
+    # L0/L4/L5 hold different (overlapping) projections of the corpus.
+    # Each layer is independently probed; transient failure on one layer
+    # leaves its slot null rather than zeroing the whole field.
+    memories: dict[str, int | None] = {
+        "l0_bm25_chunks": None,
+        "l4_vectors": None,
+        "l5_chats_chunks": None,
+        "l6_vector_chunks": None,
+        "l6_fts_chunks": None,
+    }
+    # L0 lives inside L2; L2 exposes /index-internal-stats with the counts.
+    try:
+        r = await _client().get(f"{L2_PROXY_URL}/index-internal-stats", timeout=3.0)
+        if r.status_code == 200:
+            stats = r.json()
+            memories["l0_bm25_chunks"] = int(stats.get("l0_chunks") or 0)
+    except Exception:
+        pass
+    # L4 reports n_vectors on its own /health.
+    try:
+        r = await _client().get(f"{L4_VEC_URL}/health", timeout=3.0)
+        if r.status_code == 200:
+            memories["l4_vectors"] = int(r.json().get("n_vectors") or 0)
+    except Exception:
+        pass
+    # L5 reports per-collection counts on /health. We surface chats —
+    # the only collection currently populated; emails/contacts/memory
+    # collections coming back online with L5 collection bootstrap.
+    try:
+        r = await _client().get(f"{L5_MILVUS_URL}/health", timeout=3.0)
+        if r.status_code == 200:
+            colls = r.json().get("collections") or {}
+            chats = colls.get("chats") if isinstance(colls, dict) else None
+            if isinstance(chats, dict):
+                memories["l5_chats_chunks"] = int(chats.get("count") or 0)
+    except Exception:
+        pass
+    # L6 exposes vector vs fts splits on /stats.
     try:
         r = await _client().get(f"{L6_DOC_URL}/stats", timeout=3.0)
         if r.status_code == 200:
             stats = r.json()
-            out["memories"] = max(
-                int(stats.get("vector_chunks") or 0),
-                int(stats.get("fts_chunks") or 0),
-                int(stats.get("total_chunks") or 0),
-            )
+            memories["l6_vector_chunks"] = int(stats.get("vector_chunks") or 0)
+            memories["l6_fts_chunks"] = int(stats.get("fts_chunks") or 0)
     except Exception:
-        out["memories"] = None
+        pass
+    out["memories"] = memories
+    return out
+
+
+@app.get("/health/deep")
+async def health_deep():
+    """Aggregate functional probe — fans out to each layer's /health/deep
+    and reports per-layer ok/status.
+
+    Where /health checks "process up + port answering", /health/deep
+    actually exercises embed → write → search round-trips on each layer
+    that supports it. Catches the class of bug that masquerades as
+    "healthy" — request handlers 500'ing while the process stays up.
+
+    Slower than /health (~1–2s); intended for ops/monitoring/cron use,
+    not the deploy gate or compose healthcheck.
+    """
+    import asyncio
+    out: dict[str, Any] = {"status": "ok", "ok": True, "layers": {}}
+
+    async def _probe_deep(name: str, url: str) -> tuple[str, dict]:
+        try:
+            r = await _client().get(url, timeout=15.0)
+            if r.status_code != 200:
+                return name, {"ok": False, "status": f"http {r.status_code}"}
+            return name, r.json()
+        except Exception as exc:
+            return name, {"ok": False, "status": f"unreachable: {type(exc).__name__}"}
+
+    results = await asyncio.gather(
+        _probe_deep("l4", f"{L4_VEC_URL}/health/deep"),
+        _probe_deep("l5", f"{L5_MILVUS_URL}/health/deep"),
+        _probe_deep("l6", f"{L6_DOC_URL}/health/deep"),
+    )
+    for name, body in results:
+        out["layers"][name] = body
+        if not body.get("ok", False):
+            out["ok"] = False
+
+    if not out["ok"]:
+        out["status"] = "degraded"
     return out
 
 

package/packages/memory-engine/engine/services/l2/l2-hybridrag-proxy.py

@@ -12,6 +12,7 @@ Port: 8031 (replaces neo4j-qmd-proxy.py)
 """
 
 import argparse
+import asyncio
 import hashlib
 import json
 import logging
@@ -19,14 +20,16 @@ import os
 import sqlite3
 import sys
 import time
+from contextlib import asynccontextmanager
 from datetime import datetime
 from pathlib import Path
-from typing import Any, Dict, List, Optional, Set
+from typing import Any, AsyncIterator, Dict, List, Optional, Set
 
 import re
+import httpx
 import requests
 from fastapi import FastAPI, HTTPException, Request
-from neo4j import GraphDatabase
+from neo4j import AsyncGraphDatabase
 from neo4j.time import DateTime as Neo4jDateTime, Date as Neo4jDate
 from pydantic import BaseModel
 import uvicorn
@@ -129,7 +132,67 @@ TRACKER_FILE = WORKSPACE / "memory" / "memory-tracker.jsonl"
 logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
 log = logging.getLogger("sequential-hybridrag")
 
-app = FastAPI(title="Sequential HybridRAG Proxy", version="1.0.0")
+
+# ---------------------------------------------------------------------------
+# Async driver + HTTP client singletons
+# ---------------------------------------------------------------------------
+#
+# v0.8.4: migrated from sync neo4j (`GraphDatabase`) to async
+# (`AsyncGraphDatabase`) and from per-call drivers to one process-wide
+# pool. Previously every handler created a fresh sync driver inline,
+# which (a) blocked the event loop on connection establishment and
+# every query, and (b) held one thread per in-flight request. Under
+# sustained ingest the threadpool exhausted and /health itself timed
+# out (the `l0/l1/l2: unreachable: ReadTimeout` we saw in prod).
+#
+# The async driver multiplexes many queries through bolt without
+# blocking the event loop; one shared pool means connection
+# establishment is amortised. fastapi lifespan handles open/close.
+
+_neo4j_driver: "AsyncGraphDatabase | None" = None
+_http_client: "httpx.AsyncClient | None" = None
+
+
+def get_neo4j_driver():
+    """Lazy module-level singleton — created at first call (which the
+    lifespan handler does at startup). Sharing across requests is the
+    documented neo4j Python-driver pattern; the driver itself is
+    thread- and task-safe."""
+    global _neo4j_driver
+    if _neo4j_driver is None:
+        _neo4j_driver = AsyncGraphDatabase.driver(NEO4J_URI, auth=NEO4J_AUTH)
+    return _neo4j_driver
+
+
+def get_http_client() -> httpx.AsyncClient:
+    """Shared async HTTP client for L4/L5/L6 fan-out and embedding-proxy
+    pass-through. Reuses TCP connections via httpx's built-in pool."""
+    global _http_client
+    if _http_client is None:
+        _http_client = httpx.AsyncClient(timeout=30.0)
+    return _http_client
+
+
+@asynccontextmanager
+async def lifespan(_app: FastAPI) -> AsyncIterator[None]:
+    """Open the neo4j driver + HTTP client at process startup, close on
+    shutdown. Without this, the first request pays driver-open latency
+    and the driver is never properly closed on SIGTERM (leaking conns)."""
+    global _neo4j_driver, _http_client
+    _neo4j_driver = AsyncGraphDatabase.driver(NEO4J_URI, auth=NEO4J_AUTH)
+    _http_client = httpx.AsyncClient(timeout=30.0)
+    try:
+        yield
+    finally:
+        if _neo4j_driver is not None:
+            await _neo4j_driver.close()
+            _neo4j_driver = None
+        if _http_client is not None:
+            await _http_client.aclose()
+            _http_client = None
+
+
+app = FastAPI(title="Sequential HybridRAG Proxy", version="1.0.0", lifespan=lifespan)
 
 # ---------------------------------------------------------------------------
 # Memory Usage Tracking
@@ -267,7 +330,7 @@ def extract_query_entities(query: str) -> List[str]:
     log.info(f"Extracted entities: {potential_entities}")
     return potential_entities
 
-def _hebbian_strengthen(session, arenas: List[str], node_names: List[str], increment: float = 0.05) -> None:
+async def _hebbian_strengthen(session, arenas: List[str], node_names: List[str], increment: float = 0.05) -> None:
     """Hebbian: strengthen edges between co-accessed nodes during query.
 
     Scoped by arena so a search inside tenant A can't reinforce edges
@@ -281,7 +344,7 @@ def _hebbian_strengthen(session, arenas: List[str], node_names: List[str], incre
     for i, n1 in enumerate(node_names):
         for n2 in node_names[i+1:]:
             try:
-                session.run(
+                await session.run(
                     """MATCH (a {name: $n1})-[r]-(b {name: $n2})
                        WHERE a.arena IN $arenas AND b.arena IN $arenas
                        SET r.weight = coalesce(r.weight, 1.0) + $inc,
@@ -292,7 +355,7 @@ def _hebbian_strengthen(session, arenas: List[str], node_names: List[str], incre
                 pass  # non-critical
 
 
-def search_neo4j_sequential(query: str, entities: List[str], arenas: List[str], limit: int = 12) -> Dict:
+async def search_neo4j_sequential(query: str, entities: List[str], arenas: List[str], limit: int = 12) -> Dict:
     """Phase 1: Neo4j graph search with spreading activation + Hebbian.
 
     `arenas` is the tenant-scope set the caller is authorised for —
@@ -306,11 +369,11 @@ def search_neo4j_sequential(query: str, entities: List[str], arenas: List[str],
         log.warning("search_neo4j_sequential called without arenas — returning empty results")
         return {"results": [], "graph_entities": [], "entity_count": 0}
     try:
-        driver = GraphDatabase.driver(NEO4J_URI, auth=NEO4J_AUTH)
+        driver = get_neo4j_driver()
         results = []
         graph_entities = set()
 
-        with driver.session() as session:
+        async with driver.session() as session:
             # Search for specific entities — use weighted spreading activation
             for entity in entities:
                 # Direct match first — arena-scoped on every node we touch.
@@ -326,9 +389,9 @@ def search_neo4j_sequential(query: str, entities: List[str], arenas: List[str],
                 LIMIT $limit
                 """
 
-                records = session.run(cypher, entity=entity, arenas=arenas, limit=8)
+                result = await session.run(cypher, entity=entity, arenas=arenas, limit=8)
 
-                for record in records:
+                async for record in result:
                     node = _serialize_neo4j_value(dict(record["n"]))
                     rel = record["r"]
                     connected = record["connected"]
@@ -366,7 +429,7 @@ def search_neo4j_sequential(query: str, entities: List[str], arenas: List[str],
                 # the filter, an activation could walk into another
                 # tenant's graph via a name-collision on the start node.
                 if entity:
-                    activation_results = session.run("""
+                    activation_results = await session.run("""
                         MATCH (start)-[r1]-(mid)-[r2]-(end)
                         WHERE start.name CONTAINS $entity
                           AND start.arena IN $arenas
@@ -382,7 +445,7 @@ def search_neo4j_sequential(query: str, entities: List[str], arenas: List[str],
                         LIMIT 5
                     """, entity=entity, arenas=arenas)
 
-                    for rec in activation_results:
+                    async for rec in activation_results:
                         end_node = _serialize_neo4j_value(dict(rec["end"])) if rec["end"] else {}
                         name = end_node.get("name", "")
                         if name and name not in graph_entities:
@@ -413,9 +476,9 @@ def search_neo4j_sequential(query: str, entities: List[str], arenas: List[str],
                     LIMIT $limit
                     """
 
-                    records = session.run(cypher, term=word, arenas=arenas, limit=4)
+                    result = await session.run(cypher, term=word, arenas=arenas, limit=4)
 
-                    for record in records:
+                    async for record in result:
                         node = _serialize_neo4j_value(dict(record["n"]))
                         context = f"Related: {node}"
                         graph_entities.add(node.get('name', ''))
@@ -430,9 +493,9 @@ def search_neo4j_sequential(query: str, entities: List[str], arenas: List[str],
                         })
 
             # Hebbian: strengthen edges between all accessed entities
-            _hebbian_strengthen(session, arenas, list(graph_entities))
+            await _hebbian_strengthen(session, arenas, list(graph_entities))
 
-        driver.close()
+        # Note: driver is a module-level singleton, do NOT close here.
 
         return {
             "results": results[:limit],
@@ -861,8 +924,8 @@ def search_l0_bm25(query: str, limit: int = 6, arena: str = None,
 
 L5_API_URL = os.environ.get("PME_L5_URL", "http://127.0.0.1:8034")
 
-def search_l5_communications(query: str, limit: int = 6, arena: str = None,
-                             arenas: List[str] = None) -> List[Dict]:
+async def search_l5_communications(query: str, limit: int = 6, arena: str = None,
+                                   arenas: List[str] = None) -> List[Dict]:
     """Search L5 Communications Context via L5 API (emails, chats, calendar).
 
     arena / arenas (optional): forwarded to L5; filters Milvus by the
@@ -876,7 +939,7 @@ def search_l5_communications(query: str, limit: int = 6, arena: str = None,
         params: list = [("q", query), ("limit", str(limit))]
         for a in arena_list:
             params.append(("arenas", a))
-        resp = requests.get(
+        resp = await get_http_client().get(
             f"{L5_API_URL}/search",
             params=params,
             timeout=10,
@@ -922,8 +985,8 @@ def search_l5_communications(query: str, limit: int = 6, arena: str = None,
 # L6: Document Store Search
 L6_URL = os.environ.get("PME_L6_URL", "http://localhost:8037")
 
-def search_l6_documents(query: str, limit: int = 6, arena: str = None,
-                        arenas: List[str] = None) -> List[Dict]:
+async def search_l6_documents(query: str, limit: int = 6, arena: str = None,
+                              arenas: List[str] = None) -> List[Dict]:
     """Search L6 Document Store (research, legal, financial, project docs).
 
     arena / arenas (optional): forwarded to L6 — L6 supports multi-arena
@@ -939,7 +1002,7 @@ def search_l6_documents(query: str, limit: int = 6, arena: str = None,
         ]
         for a in arena_list:
             params.append(("arenas", a))
-        resp = requests.get(
+        resp = await get_http_client().get(
             f"{L6_URL}/search",
             params=params,
             timeout=10,
@@ -986,50 +1049,52 @@ def search_l6_documents(query: str, limit: int = 6, arena: str = None,
         return []
 
 
-def sequential_hybridrag_search(query: str, limit: int = 16,
-                                arena: str = None,
-                                arenas: List[str] = None) -> List[Dict]:
+async def sequential_hybridrag_search(query: str, limit: int = 16,
+                                      arena: str = None,
+                                      arenas: List[str] = None) -> List[Dict]:
     """Main HybridRAG processing: L0 BM25 → L1 System Files → L2 HybridRAG (L3 Graph + L4 Vector + L5 Comms + L6 Docs).
 
     arena / arenas (optional): tenant + user scope. Multi-arena lets a
     user's search span tenant-wide rows + their own user-scoped rows in
     a single hybrid pass. Forwarded to L0, L5, L6 native filters; L4
     and L3 still rely on the compat shim post-filter.
+
+    Async since v0.8.4: independent layer fan-out runs concurrently via
+    asyncio.gather; sync workers (sqlite, file I/O, PyTorch reranker)
+    are dispatched to threads via asyncio.to_thread to keep the event
+    loop responsive under sustained ingest.
     """
     arena_list = list(arenas) if arenas else ([arena] if arena else [])
     start_time = time.time()
     log.info(f"Starting sequential HybridRAG search for: '{query}' arenas={arena_list!r}")
 
-    # L0: BM25 workspace memory (keyword search — complements semantic layers)
-    l0_results = search_l0_bm25(query, limit=6, arenas=arena_list)
+    # L0 (sqlite) + L1 (file I/O) + entity extraction run in parallel.
+    # Each is sync — offloaded to the threadpool so the event loop stays
+    # free for the in-flight L3/L5/L6 calls.
+    l0_results, system_results, entities = await asyncio.gather(
+        asyncio.to_thread(search_l0_bm25, query, 6, None, arena_list),
+        asyncio.to_thread(search_core_memory_files, query, 4),
+        asyncio.to_thread(extract_query_entities, query),
+    )
     log.info(f"L0 BM25 workspace: {len(l0_results)} results")
-
-    # L1: System Files (HIGHEST PRIORITY)
-    system_results = search_core_memory_files(query, limit=4)
     log.info(f"L1 System files: {len(system_results)} results")
 
-    # L2: HybridRAG orchestration
-    # L3: Graph search (entity extraction + Neo4j) — arena-scoped so a
-    # tenant's search can never traverse another tenant's entity graph
-    # via name collisions on shared :Entity nodes. The post-filter shim
-    # protects chunks; this protects the entity-walking layer too.
-    entities = extract_query_entities(query)
-    graph_context = search_neo4j_sequential(query, entities, arena_list, limit=8)
+    # L3: Graph search (now native async via AsyncGraphDatabase).
+    graph_context = await search_neo4j_sequential(query, entities, arena_list, limit=8)
     log.info(f"L3 Graph search: {len(graph_context['results'])} results, {graph_context['entity_count']} entities")
 
-    # HyDE: expand query for better vector embeddings
-    hyde_query = hyde_expand(query)
+    # HyDE expansion is sync (LLM call), offload to thread.
+    hyde_query = await asyncio.to_thread(hyde_expand, query)
 
-    # L4: Vector search (informed by L3 graph context + HyDE)
-    vector_results = search_qmd_informed(hyde_query, graph_context, limit=8)
+    # L4/L5/L6 fan out concurrently — L4 (sqlite) via to_thread, L5/L6
+    # native async via httpx.AsyncClient.
+    vector_results, l5_results, l6_results = await asyncio.gather(
+        asyncio.to_thread(search_qmd_informed, hyde_query, graph_context, 8),
+        search_l5_communications(hyde_query, limit=6, arenas=arena_list),
+        search_l6_documents(hyde_query, limit=6, arenas=arena_list),
+    )
     log.info(f"L4 Vector search: {len(vector_results)} results (HyDE={'on' if hyde_query != query else 'off'})")
-
-    # L5: Communications Context (emails, chats, calendar) — also use HyDE
-    l5_results = search_l5_communications(hyde_query, limit=6, arenas=arena_list)
     log.info(f"L5 Communications: {len(l5_results)} results")
-
-    # L6: Document Store (research, legal, financial, project docs)
-    l6_results = search_l6_documents(hyde_query, limit=6, arenas=arena_list)
     log.info(f"L6 Documents: {len(l6_results)} results")
 
     # L2: HybridRAG fusion (combines all layers with L1 priority)
@@ -1046,8 +1111,9 @@ def sequential_hybridrag_search(query: str, limit: int = 16,
     # Sort by layer priority: L1 System (1.0) > L3 Graph (0.9) > L4 Vector (0.7+)
     deduplicated.sort(key=lambda x: x["score"], reverse=True)
 
-    # Cross-encoder reranking: re-embed top results and blend scores
-    deduplicated = cross_encoder_rerank(query, deduplicated, top_k=limit)
+    # Cross-encoder reranking: re-embed top results and blend scores.
+    # PyTorch CrossEncoder.predict is sync — offload to thread.
+    deduplicated = await asyncio.to_thread(cross_encoder_rerank, query, deduplicated, limit)
 
     # Track layer usage for evolution
     search_time_ms = (time.time() - start_time) * 1000
@@ -1094,7 +1160,7 @@ async def search_endpoint(request: Request) -> dict:
         if not query:
             raise HTTPException(status_code=400, detail="query is required")
 
-        results = sequential_hybridrag_search(query, limit=limit, arena=arena, arenas=arenas)
+        results = await sequential_hybridrag_search(query, limit=limit, arena=arena, arenas=arenas)
 
         # Also return raw graph entities for context enrichment.
         # Same arena scope as the cascade search above — without it
@@ -1172,7 +1238,7 @@ async def chat_completions(request: ChatCompletionRequest) -> dict:
         # empty when no arenas are supplied; callers that need L3 must
         # pass `arena` or `arenas` on the request body.
         start_time = time.time()
-        results = sequential_hybridrag_search(
+        results = await sequential_hybridrag_search(
             query, limit=16, arena=request.arena, arenas=request.arenas,
         )
         search_time = time.time() - start_time
@@ -1244,33 +1310,35 @@ async def check_contradictions(node_name: str, arena: Optional[str] = None) -> d
             detail="arena query parameter is required to scope contradiction lookup",
         )
     try:
-        driver = GraphDatabase.driver(NEO4J_URI, auth=NEO4J_AUTH)
+        driver = get_neo4j_driver()
         contradictions = []
-        with driver.session() as session:
+        async with driver.session() as session:
             # Find the node — must be in the caller's arena.
-            node = session.run(
+            result = await session.run(
                 """MATCH (n) WHERE toLower(n.name) = toLower($name) AND n.arena = $arena
                    RETURN elementId(n) AS id""",
                 name=node_name, arena=arena,
-            ).single()
+            )
+            node = await result.single()
             if not node:
                 return {"node": node_name, "contradictions": [], "error": "Node not found"}
             nid = node["id"]
 
             # Explicit CONTRADICTS — both endpoints must be in the same arena.
-            for rec in session.run(
+            result = await session.run(
                 """MATCH (a)-[r:CONTRADICTS]-(b)
                    WHERE elementId(a) = $nid AND b.arena = $arena
                    RETURN a.name AS a, b.name AS b, r.reason AS reason""",
                 nid=nid, arena=arena,
-            ):
+            )
+            async for rec in result:
                 contradictions.append({"type": "explicit", "a": rec["a"], "b": rec["b"], "reason": rec["reason"]})
 
             # Property conflicts via shared neighbour — every node along
             # the (a)--(shared)--(b) path filtered by arena so a shared
             # neighbour from another tenant can't trigger a false-positive
             # conflict in this tenant's view.
-            for rec in session.run(
+            result = await session.run(
                 """MATCH (a)--(shared)--(b)
                    WHERE elementId(a) = $nid AND a <> b
                      AND shared.arena = $arena AND b.arena = $arena
@@ -1281,28 +1349,28 @@ async def check_contradictions(node_name: str, arena: Optional[str] = None) -> d
                    WHERE size(ck) > 0
                    RETURN a.name AS a, b.name AS b, shared.name AS via, ck
                    LIMIT 10""", nid=nid, arena=arena,
-            ):
+            )
+            async for rec in result:
                 contradictions.append({
                     "type": "property_conflict", "a": rec["a"], "b": rec["b"],
                     "via": rec["via"], "conflicting_keys": rec["ck"]
                 })
-        driver.close()
         return {"node": node_name, "contradictions": contradictions, "count": len(contradictions)}
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))
 
-def _check_l5_health() -> bool:
+async def _check_l5_health() -> bool:
     """Quick check if L5 Communications API is responding."""
     try:
-        resp = requests.get(f"{L5_API_URL}/health", timeout=3)
+        resp = await get_http_client().get(f"{L5_API_URL}/health", timeout=3)
         return resp.status_code == 200
     except Exception:
         return False
 
-def _check_l6_health() -> bool:
+async def _check_l6_health() -> bool:
     """Quick check if L6 Document Store is responding."""
     try:
-        resp = requests.get(f"{L6_URL}/health", timeout=3)
+        resp = await get_http_client().get(f"{L6_URL}/health", timeout=3)
         return resp.status_code == 200 and resp.json().get("status") in ("ok", "degraded")
     except Exception:
         return False
@@ -1323,16 +1391,15 @@ async def health() -> dict:
 
     neo4j_healthy = False
     try:
-        driver = GraphDatabase.driver(NEO4J_URI, auth=NEO4J_AUTH)
-        with driver.session() as session:
-            session.run("RETURN 1")
+        driver = get_neo4j_driver()
+        async with driver.session() as session:
+            await session.run("RETURN 1")
         neo4j_healthy = True
-        driver.close()
     except Exception as e:
         logging.debug(f"Suppressed: {e}")
 
-    l5_reachable = _check_l5_health()
-    l6_reachable = _check_l6_health()
+    l5_reachable = await _check_l5_health()
+    l6_reachable = await _check_l6_health()
 
     # Top-level status: degrade only on layers L2 is the sole gatekeeper for.
     # L5/L6 are independent services probed by the compat shim.
@@ -1411,15 +1478,19 @@ def _extract_entities_for_kg(text: str, max_entities: int = 32) -> List[str]:
     return found[:max_entities]
 
 
-def _embed_batch_local(texts: List[str]) -> List[List[float]]:
-    """Batch embed via the shared EmbedClient. Returns vectors in input order."""
+async def _embed_batch_local(texts: List[str]) -> List[List[float]]:
+    """Batch embed via the shared EmbedClient (async). Returns vectors in input order."""
     if not texts:
         return []
     try:
-        return _embed_client().embed_batch(texts)
+        return await _embed_client().embed_batch_async(texts)
     except Exception as e:
         log.warning(f"NV-Embed batch failed: {e}; trying singletons")
-        return [get_embedding(t) for t in texts]
+        # Singleton fallback stays sync (each one-shot embed is small);
+        # offload to thread so we don't block the loop.
+        return await asyncio.gather(
+            *(asyncio.to_thread(get_embedding, t) for t in texts)
+        )
 
 
 class IndexInternalBatchRequest(BaseModel):
@@ -1504,7 +1575,7 @@ async def index_internal_batch(req: IndexInternalBatchRequest) -> dict:
     # ---- L4 QMD vec (qmd.sqlite) ----------------------------------------
     l4_inserted = 0
     try:
-        embeddings = _embed_batch_local([n["content"] for n in norm])
+        embeddings = await _embed_batch_local([n["content"] for n in norm])
         if len(embeddings) != len(norm):
             log.warning(f"L4 embed count mismatch: {len(embeddings)} != {len(norm)}")
         qmd_db = Path(QMD_DB_PATH)
@@ -1560,30 +1631,30 @@ async def index_internal_batch(req: IndexInternalBatchRequest) -> dict:
     l3_entities = 0
     l3_chunks = 0
     try:
-        driver = GraphDatabase.driver(NEO4J_URI, auth=NEO4J_AUTH)
-        with driver.session() as session:
+        driver = get_neo4j_driver()
+        async with driver.session() as session:
             # Indexes — idempotent. The compound (arena, name) is the
             # right shape now that entities are arena-scoped; the legacy
             # entity_name index stays for the wipe-migration to work
             # against pre-arena rows, then can be dropped in a follow-up.
             try:
-                session.run("CREATE INDEX entity_arena_name IF NOT EXISTS FOR (n:Entity) ON (n.arena, n.name)")
-                session.run("CREATE INDEX person_arena_email IF NOT EXISTS FOR (n:Person) ON (n.arena, n.email)")
-                session.run("CREATE INDEX chunk_arena IF NOT EXISTS FOR (c:Chunk) ON (c.arena)")
-                session.run("CREATE INDEX chunk_id IF NOT EXISTS FOR (c:Chunk) ON (c.id)")
+                await session.run("CREATE INDEX entity_arena_name IF NOT EXISTS FOR (n:Entity) ON (n.arena, n.name)")
+                await session.run("CREATE INDEX person_arena_email IF NOT EXISTS FOR (n:Person) ON (n.arena, n.email)")
+                await session.run("CREATE INDEX chunk_arena IF NOT EXISTS FOR (c:Chunk) ON (c.arena)")
+                await session.run("CREATE INDEX chunk_id IF NOT EXISTS FOR (c:Chunk) ON (c.id)")
                 # ChannelStat is the denormalised aggregate read by
                 # /aggregate on the fast path. Compound index covers
                 # the (arena, person_email) lookup that the reader
                 # uses; the per-channel rows are returned in one
                 # range scan.
-                session.run("CREATE INDEX channelstat_arena_email IF NOT EXISTS FOR (s:ChannelStat) ON (s.arena, s.person_email)")
+                await session.run("CREATE INDEX channelstat_arena_email IF NOT EXISTS FOR (s:ChannelStat) ON (s.arena, s.person_email)")
                 # UNIQUE constraint on the writer's MERGE key. Without
                 # this, two concurrent index-internal-batch transactions
                 # can both decide a ChannelStat doesn't exist and create
                 # rival nodes — the index doesn't lock, the constraint
                 # does. The constraint also implies an index on the
                 # full key so the MERGE locks efficiently.
-                session.run("CREATE CONSTRAINT channelstat_unique IF NOT EXISTS FOR (s:ChannelStat) REQUIRE (s.arena, s.person_email, s.channel) IS UNIQUE")
+                await session.run("CREATE CONSTRAINT channelstat_unique IF NOT EXISTS FOR (s:ChannelStat) REQUIRE (s.arena, s.person_email, s.channel) IS UNIQUE")
             except Exception:
                 pass
             for n in norm:
@@ -1602,7 +1673,7 @@ async def index_internal_batch(req: IndexInternalBatchRequest) -> dict:
                 # tenant-isolation anchor. Every read traverses through
                 # this node, so getting the arena right here is the
                 # single most important invariant of this whole block.
-                session.run(
+                await session.run(
                     """
                     MERGE (c:Chunk {id: $cid})
                     SET c.text = $text,
@@ -1618,7 +1689,7 @@ async def index_internal_batch(req: IndexInternalBatchRequest) -> dict:
 
                 # Concept entities — heuristic, arena-scoped.
                 for ent in heuristic_entities:
-                    session.run(
+                    await session.run(
                         """
                         MERGE (e:Entity:Concept {arena: $arena, name: $name})
                         ON CREATE SET e.type = 'Concept',
@@ -1639,7 +1710,7 @@ async def index_internal_batch(req: IndexInternalBatchRequest) -> dict:
                 if len(heuristic_entities) >= 2:
                     for i in range(len(heuristic_entities)):
                         for j in range(i + 1, len(heuristic_entities)):
-                            session.run(
+                            await session.run(
                                 """
                                 MATCH (a:Entity:Concept {arena: $arena, name: $a})
                                 MATCH (b:Entity:Concept {arena: $arena, name: $b})
@@ -1666,7 +1737,7 @@ async def index_internal_batch(req: IndexInternalBatchRequest) -> dict:
                     # flipped to true after the stat update, so replays
                     # of the same eventId never double-count even when
                     # the chunk already exists.
-                    session.run(
+                    await session.run(
                         """
                         MERGE (p:Entity:Person {arena: $arena, email: $email})
                         ON CREATE SET p.created_at = $now,
@@ -1725,7 +1796,7 @@ async def index_internal_batch(req: IndexInternalBatchRequest) -> dict:
                     l3_entities += 1
                 if isinstance(contact_name, str) and contact_name.strip():
                     cname = contact_name.strip()
-                    session.run(
+                    await session.run(
                         """
                         MERGE (p:Entity:Person {arena: $arena, name: $name})
                         ON CREATE SET p.created_at = $now,
@@ -1751,7 +1822,7 @@ async def index_internal_batch(req: IndexInternalBatchRequest) -> dict:
                     # Link name→email node so the relationships query
                     # can resolve either alias to the same person.
                     if person_email_node:
-                        session.run(
+                        await session.run(
                             """
                             MATCH (n:Person {arena: $arena, name: $name})
                             MATCH (e:Person {arena: $arena, email: $email})
@@ -1759,7 +1830,6 @@ async def index_internal_batch(req: IndexInternalBatchRequest) -> dict:
                             """,
                             arena=arena, name=cname, email=person_email_node,
                         )
-        driver.close()
     except Exception as e:
         log.error(f"L3 KG write failed: {e}")
 
@@ -1841,25 +1911,28 @@ async def forget_internal(request: Request) -> dict:
     # Neo4j chunks AND entities both carry arena now, so tenant-scoped
     # delete works correctly here even if L0/L4 still need a migration.
     try:
-        driver = GraphDatabase.driver(NEO4J_URI, auth=NEO4J_AUTH)
-        with driver.session() as session:
+        driver = get_neo4j_driver()
+        async with driver.session() as session:
             if arena:
-                r1 = session.run(
+                r1 = await session.run(
                     "MATCH (c:Chunk {arena: $arena}) DETACH DELETE c RETURN count(c) AS n",
                     arena=arena,
                 )
-                deleted["l3_chunks"] = r1.single()["n"]
-                r2 = session.run(
+                rec = await r1.single()
+                deleted["l3_chunks"] = rec["n"]
+                r2 = await session.run(
                     "MATCH (e:Entity {arena: $arena}) DETACH DELETE e RETURN count(e) AS n",
                     arena=arena,
                 )
-                deleted["l3_entities"] = r2.single()["n"]
+                rec = await r2.single()
+                deleted["l3_entities"] = rec["n"]
             else:  # confirm == "GLOBAL_WIPE", validated above
-                r1 = session.run("MATCH (c:Chunk) DETACH DELETE c RETURN count(c) AS n")
-                deleted["l3_chunks"] = r1.single()["n"]
-                r2 = session.run("MATCH (e:Entity) DETACH DELETE e RETURN count(e) AS n")
-                deleted["l3_entities"] = r2.single()["n"]
-        driver.close()
+                r1 = await session.run("MATCH (c:Chunk) DETACH DELETE c RETURN count(c) AS n")
+                rec = await r1.single()
+                deleted["l3_chunks"] = rec["n"]
+                r2 = await session.run("MATCH (e:Entity) DETACH DELETE e RETURN count(e) AS n")
+                rec = await r2.single()
+                deleted["l3_entities"] = rec["n"]
     except Exception as e:
         log.error(f"L3 forget failed: {e}")
     return {"status": "ok", "deleted": deleted, "arena": arena, "global_wipe": confirm == "GLOBAL_WIPE"}
@@ -1948,13 +2021,10 @@ async def aggregate_internal(req: AggregateInternalRequest) -> AggregateInternal
             seen.add(k)
             safe_group_by.append(k)
 
-    try:
-        driver = GraphDatabase.driver(NEO4J_URI, auth=NEO4J_AUTH)
-    except Exception as e:
-        raise HTTPException(status_code=500, detail=f"neo4j connect: {e}")
+    driver = get_neo4j_driver()
 
     try:
-        with driver.session() as session:
+        async with driver.session() as session:
             # Fast path: read from the ChannelStat denormalisation
             # whenever the caller has an email and is grouping by
             # channel. ChannelStats are written by /index-internal-batch
@@ -1973,7 +2043,7 @@ async def aggregate_internal(req: AggregateInternalRequest) -> AggregateInternal
                 not safe_group_by or safe_group_by == ["channel"]
             )
             if fast_path_eligible:
-                stats_rows = list(session.run(
+                _res = await session.run(
                     "MATCH (s:ChannelStat {arena: $arena, person_email: $email})\n"
                     "RETURN s.channel AS channel,\n"
                     "       s.count AS count,\n"
@@ -1983,7 +2053,8 @@ async def aggregate_internal(req: AggregateInternalRequest) -> AggregateInternal
                     "       s.first_seen AS first_seen\n"
                     "ORDER BY s.count DESC\n",
                     arena=arena, email=contact_email,
-                ))
+                )
+                stats_rows = [rec async for rec in _res]
                 if stats_rows:
                     # Build buckets directly. When group_by=[] we
                     # collapse to a single overall bucket; otherwise
@@ -2098,7 +2169,8 @@ async def aggregate_internal(req: AggregateInternalRequest) -> AggregateInternal
             buckets: List[AggregateBucket] = []
             total = 0
             latest: Optional[str] = None
-            for rec in session.run(cypher, **params):
+            _res = await session.run(cypher, **params)
+            async for rec in _res:
                 count = int(rec["count"] or 0)
                 total += count
                 last_seen = rec["last_seen"]
@@ -2126,8 +2198,6 @@ async def aggregate_internal(req: AggregateInternalRequest) -> AggregateInternal
     except Exception as e:
         log.error(f"aggregate-internal failed: {e}")
         raise HTTPException(status_code=500, detail=f"aggregate failed: {e}")
-    finally:
-        driver.close()
 
 
 @app.get("/index-internal-stats")
@@ -2153,13 +2223,14 @@ async def index_internal_stats() -> dict:
     except Exception as e:
         out["l4_qmd_error"] = str(e)
     try:
-        driver = GraphDatabase.driver(NEO4J_URI, auth=NEO4J_AUTH)
-        with driver.session() as session:
-            r = session.run("MATCH (c:Chunk) RETURN count(c) AS n").single()
+        driver = get_neo4j_driver()
+        async with driver.session() as session:
+            res = await session.run("MATCH (c:Chunk) RETURN count(c) AS n")
+            r = await res.single()
             out["l3_chunks"] = r["n"] if r else 0
-            r = session.run("MATCH (e:Entity) RETURN count(e) AS n").single()
+            res = await session.run("MATCH (e:Entity) RETURN count(e) AS n")
+            r = await res.single()
             out["l3_entities"] = r["n"] if r else 0
-        driver.close()
     except Exception as e:
         out["l3_error"] = str(e)
     return out

package/packages/memory-engine/engine/services/l4/server.py

@@ -136,7 +136,11 @@ def health():
         n = conn.execute("SELECT COUNT(*) FROM chunks").fetchone()[0]
         conn.close()
         return {"status": "ok", "loaded": True, "n_vectors": n,
-                "dim": EMBED_DIM, "db_path": DB_PATH, "backend": "sqlite-vec-fallback"}
+                "dim": EMBED_DIM, "db_path": DB_PATH,
+                # BLOB+Python-cosine is the intentional implementation path,
+                # not a degraded fallback (see _get_db docstring). The previous
+                # "sqlite-vec-fallback" label gave operators the wrong signal.
+                "backend": "sqlite-vec"}
     except Exception as exc:
         return {"status": "degraded", "error": str(exc)}
 
@@ -212,6 +216,80 @@ def refresh():
     return {"status": "ok", "noop": True}
 
 
+# ----------------------------------------------------------------------
+# /health/deep — synthetic round-trip
+# ----------------------------------------------------------------------
+
+# Fixed sentinel id used by /health/deep. Upserted on every probe call,
+# so the row is idempotent. Kept under id="__healthcheck__sentinel" so
+# the L4 corpus has at most one healthcheck row regardless of probe rate.
+_HEALTH_SENTINEL_ID = "__healthcheck__sentinel"
+_HEALTH_SENTINEL_TEXT = (
+    "healthcheck sentinel — embed-write-search round-trip verifier"
+)
+
+
+@app.get("/health/deep")
+async def health_deep():
+    """Real functional probe: embed → write → search the sentinel.
+
+    Catches the class of failure that plain /health misses — broken
+    embed paths, write 500s, query path bugs — i.e. exactly the bug
+    shape that silently degraded L6 from v0.8.0 → v0.8.2.
+
+    Returns:
+        {status, embed_ms, write_ms, search_ms, hit, ok}
+
+    `hit` confirms the sentinel was returned from search; `ok` is the
+    aggregate `embed_ok AND write_ok AND hit`. HTTP status stays 200
+    regardless so callers can read the body for diagnostics; status:
+    field carries the verdict.
+    """
+    t_total = time.perf_counter()
+    out: dict[str, Any] = {"status": "ok", "ok": True}
+    try:
+        t0 = time.perf_counter()
+        embs = await _embed_batch([_HEALTH_SENTINEL_TEXT])
+        out["embed_ms"] = round((time.perf_counter() - t0) * 1000.0, 1)
+        if not embs or not embs[0]:
+            out["status"] = "embed_failed"
+            out["ok"] = False
+            return out
+        vec = embs[0]
+    except Exception as exc:
+        out["status"] = f"embed_error: {type(exc).__name__}"
+        out["ok"] = False
+        return out
+
+    try:
+        conn = _get_db()
+        t1 = time.perf_counter()
+        conn.execute(
+            "INSERT OR REPLACE INTO chunks(id, text, embedding, indexed_at) "
+            "VALUES (?, ?, ?, ?)",
+            (_HEALTH_SENTINEL_ID, _HEALTH_SENTINEL_TEXT, _vec_to_blob(vec), time.time()),
+        )
+        conn.commit()
+        out["write_ms"] = round((time.perf_counter() - t1) * 1000.0, 1)
+
+        t2 = time.perf_counter()
+        rows = conn.execute("SELECT id, embedding FROM chunks WHERE id = ?",
+                            (_HEALTH_SENTINEL_ID,)).fetchone()
+        out["search_ms"] = round((time.perf_counter() - t2) * 1000.0, 1)
+        conn.close()
+    except Exception as exc:
+        out["status"] = f"db_error: {type(exc).__name__}"
+        out["ok"] = False
+        return out
+
+    out["hit"] = rows is not None
+    if not out["hit"]:
+        out["status"] = "sentinel_missing"
+        out["ok"] = False
+    out["total_ms"] = round((time.perf_counter() - t_total) * 1000.0, 1)
+    return out
+
+
 # ----------------------------------------------------------------------
 # Entrypoint
 # ----------------------------------------------------------------------

package/packages/memory-engine/engine/services/l5/l5-comms-layer.py

@@ -558,6 +558,20 @@ def serve(port=8034):
     from fastapi import FastAPI, Query
     import uvicorn
 
+    # Bootstrap all 4 collections on startup. Previously only `chats`
+    # was being created (the indexer entrypoints below each call their
+    # own ensure_collection lazily, so collections without an indexer
+    # — i.e. those fed solely via /index-internal or the compat shim's
+    # _index_l5 fan-out — never came into existence and writes to them
+    # 500'd). Idempotent: ensure_collection short-circuits if exists.
+    try:
+        bootstrap_client = get_client()
+        for _name in ("chats", "emails", "contacts", "memory"):
+            ensure_collection(bootstrap_client, _name)
+        logging.info("L5 collections bootstrapped: chats, emails, contacts, memory")
+    except Exception as exc:
+        logging.warning(f"L5 collection bootstrap failed (continuing): {exc}")
+
     api = FastAPI(title="L5 Communications Layer")
 
     @api.get("/health")
@@ -658,6 +672,71 @@ def serve(port=8034):
             "insert_ms": round(insert_ms, 1),
         }
 
+    @api.get("/health/deep")
+    def api_health_deep():
+        """Real functional probe: assert all 4 collections exist + run
+        embed+insert+search of a sentinel chunk in the `chats` collection.
+        Catches missing-collection regressions (Issue 3) and embed/insert/
+        search path bugs the shallow /health misses."""
+        import time as _time, hashlib as _hashlib
+        out = {"status": "ok", "ok": True}
+        client = get_client()
+
+        # 1. Collection presence
+        expected = ("chats", "emails", "contacts", "memory")
+        present = {n: client.has_collection(n) for n in expected}
+        out["collections"] = present
+        missing = [n for n, ok in present.items() if not ok]
+        if missing:
+            out["status"] = f"missing_collections:{','.join(missing)}"
+            out["ok"] = False
+            return out
+
+        # 2. Synthetic embed → insert → search in `chats`
+        sentinel_id = "__healthcheck__sentinel"
+        sentinel_text = "healthcheck sentinel — L5 embed-write-search round-trip verifier"
+        try:
+            t0 = _time.time()
+            embs = _embed_post([sentinel_text])
+            out["embed_ms"] = round((_time.time() - t0) * 1000.0, 1)
+            if not embs or embs[0] is None:
+                out["status"] = "embed_failed"
+                out["ok"] = False
+                return out
+
+            t1 = _time.time()
+            from datetime import datetime as _dt, timezone as _tz
+            client.upsert(collection_name="chats", data=[{
+                "id": sentinel_id,
+                "vector": embs[0],
+                "text": sentinel_text,
+                "source": "healthcheck",
+                "channel": "__healthcheck__",
+                "contact": "",
+                "timestamp": _dt.now(_tz.utc).strftime("%Y-%m-%dT%H:%M:%SZ"),
+                "arena": "__healthcheck__",
+            }])
+            out["write_ms"] = round((_time.time() - t1) * 1000.0, 1)
+
+            t2 = _time.time()
+            hits = client.search(
+                collection_name="chats",
+                data=[embs[0]],
+                limit=1,
+                filter='arena == "__healthcheck__"',
+                output_fields=["id"],
+            )
+            out["search_ms"] = round((_time.time() - t2) * 1000.0, 1)
+            found = bool(hits and hits[0] and hits[0][0].get("entity", {}).get("id") == sentinel_id)
+            out["hit"] = found
+            if not found:
+                out["status"] = "sentinel_missing"
+                out["ok"] = False
+        except Exception as exc:
+            out["status"] = f"probe_error: {type(exc).__name__}: {exc}"
+            out["ok"] = False
+        return out
+
     print(f"\n  L5 Communications Layer — http://127.0.0.1:{port}")
     uvicorn.run(api, host=os.environ.get("HOST","127.0.0.1"), port=port, log_level="warning")
 

package/packages/memory-engine/engine/services/l6/Dockerfile

@@ -1,8 +1,25 @@
 FROM python:3.12-slim
 WORKDIR /app
-RUN apt-get update && apt-get install -y curl && rm -rf /var/lib/apt/lists/*
-RUN pip install --no-cache-dir fastapi "uvicorn[standard]" httpx "pymilvus[milvus_lite]" "setuptools<70" pydantic spacy
+# gcc/g++ needed by some sentence-transformers transitive deps; curl kept
+# for in-container debugging.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        gcc g++ curl \
+    && rm -rf /var/lib/apt/lists/*
+# Core deps + sentence-transformers/torch for the cross-encoder reranker.
+# Torch CPU wheel is enough — the reranker is small (MiniLM L-6) and
+# CPU-bound throughput is fine at L6's request volume. Without these,
+# get_reranker() falls back to RRF-only, capping recall ranking quality.
+RUN pip install --no-cache-dir \
+        fastapi "uvicorn[standard]" httpx "pymilvus[milvus_lite]" "setuptools<70" pydantic spacy \
+        "sentence-transformers" \
+        "torch" --extra-index-url https://download.pytorch.org/whl/cpu
 RUN python -m spacy download en_core_web_sm
+ENV HF_HOME=/data/.cache/huggingface
+# Pre-download the cross-encoder so cold-start doesn't pay first-pull
+# latency. The model is small (~80MB) and gets cached at /data — survives
+# container recreates since /data is a volume mount.
+RUN mkdir -p /data/.cache/huggingface && \
+    python -c "from sentence_transformers import CrossEncoder; CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2', cache_folder='/data/.cache/huggingface')"
 # Shared embed_provider module (build context is engine/services).
 COPY _shared /app/_shared
 COPY l6/l6-document-store.py /app/server.py

package/packages/memory-engine/engine/services/l6/l6-document-store.py

@@ -838,6 +838,67 @@ def serve(port: int = DEFAULT_PORT):
     def api_health():
         return health()
 
+    @api.get("/health/deep")
+    def api_health_deep():
+        """Real functional probe: embed → insert via /index-batch path →
+        search the sentinel via hybrid search → assert reranker loaded.
+
+        Built to catch the v0.8.0–0.8.2 L6 _embed_client shadowing bug
+        and its kind (request-handler-level breakage with the layer
+        process appearing healthy)."""
+        import time as _time
+        sentinel_id = "__healthcheck__sentinel"
+        sentinel_text = "healthcheck sentinel — L6 embed-write-search round-trip verifier"
+        out = {"status": "ok", "ok": True}
+        try:
+            t0 = _time.time()
+            try:
+                emb = embed_text(sentinel_text)
+            except Exception as exc:
+                out["status"] = f"embed_failed: {type(exc).__name__}: {exc}"
+                out["ok"] = False
+                return out
+            out["embed_ms"] = round((_time.time() - t0) * 1000.0, 1)
+
+            # Insert via the same path real ingest uses, so the probe
+            # actually exercises /index-batch's code.
+            t1 = _time.time()
+            import httpx as _httpx
+            r = _httpx.post(
+                f"http://localhost:{DEFAULT_PORT}/index-batch",
+                json={
+                    "arena": "__healthcheck__",
+                    "records": [{"id": sentinel_id, "text": sentinel_text}],
+                },
+                timeout=15.0,
+            )
+            out["write_ms"] = round((_time.time() - t1) * 1000.0, 1)
+            if r.status_code != 200:
+                out["status"] = f"write_failed: http {r.status_code}"
+                out["ok"] = False
+                return out
+
+            t2 = _time.time()
+            results = search(
+                sentinel_text, method="hybrid", limit=3,
+                arena="__healthcheck__", enable_rerank=False,
+            )
+            out["search_ms"] = round((_time.time() - t2) * 1000.0, 1)
+            hit = any(r.get("id", "").startswith(sentinel_id) for r in (results or []))
+            out["hit"] = hit
+            if not hit:
+                out["status"] = "sentinel_missing"
+                out["ok"] = False
+
+            # Reranker check — informational. Failure here doesn't flip
+            # ok=False because L6 falls back to RRF and still serves
+            # results; it just caps the recall ranking quality.
+            out["reranker"] = "ok" if get_reranker() is not None else "rrf_fallback"
+        except Exception as exc:
+            out["status"] = f"probe_error: {type(exc).__name__}: {exc}"
+            out["ok"] = False
+        return out
+
     @api.get("/stats")
     def api_stats():
         return get_stats()