@pentatonic-ai/ai-agent-sdk 0.8.1 → 0.8.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pentatonic-ai/ai-agent-sdk",
-  "version": "0.8.1",
+  "version": "0.8.3",
   "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

@@ -416,8 +416,14 @@ async def health():
         "engine": "pentatonic-memory-engine",
         "layers": {},
     }
-    # NV-Embed exposes /health alongside /v1/embeddings.
-    nv_embed_health = NV_EMBED_URL.replace("/v1/embeddings", "/health")
+    # NV-Embed (or the upstream gateway) exposes /health at the host root.
+    # Use urlparse so we swap *just* the path component instead of doing a
+    # string replace — that breaks the moment NV_EMBED_URL is /v1/embed,
+    # /v1/embeddings, or bare-host. The probe is informational only; gateways
+    # that return non-200 on root-/health don't block engine operation.
+    from urllib.parse import urlparse, urlunparse
+    _u = urlparse(NV_EMBED_URL)
+    nv_embed_health = urlunparse((_u.scheme, _u.netloc, "/health", "", "", ""))
 
     import asyncio
     l2_v, l4_v, l5_v, l6_v, nv_v, l3_v = await asyncio.gather(

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

@@ -86,7 +86,12 @@ log = logging.getLogger("l6-document-store")
 # Embedding
 # ---------------------------------------------------------------------------
 
-_embed_client = httpx.Client(timeout=60)
+# HTTP client for Ollama entity extraction (extract_entities below). Named
+# `_ollama_http`, not `_embed_client`, because the embedding HTTP client now
+# lives behind the EmbedClient factory above — sharing the `_embed_client`
+# identifier caused a TypeError in v0.8.0–0.8.2 where the legacy module-level
+# binding shadowed the factory function.
+_ollama_http = httpx.Client(timeout=60)
 
 def embed_text(text: str) -> List[float]:
     """Single-text embed via _embed_post (OpenAI-compat first, lambda-gateway fallback)."""
@@ -141,7 +146,7 @@ def rerank(query: str, results: List[Dict], top_k: int = 10) -> List[Dict]:
 def extract_entities(text: str) -> List[str]:
     """Extract entities from text using Ollama graph-preflexor."""
     try:
-        resp = _embed_client.post(
+        resp = _ollama_http.post(
             f"{OLLAMA_URL}/api/generate",
             json={
                 "model": "graph-preflexor",

package/packages/memory-engine/tests/test_compat_nv_embed_probe.py

@@ -0,0 +1,48 @@
+"""Regression test for the nv_embed /health probe URL construction.
+
+Before v0.8.1 the probe was built via str.replace("/v1/embeddings", "/health"),
+which silently produced the wrong URL when NV_EMBED_URL was anything other
+than ".../v1/embeddings" (e.g. /v1/embed for the Pentatonic AI Gateway, or
+a bare host). The probe would end up GET-ing the embed endpoint itself and
+the gateway would return 401 — making /health falsely report a degraded
+nv_embed layer even when embeddings worked fine.
+
+The fix replaces the substring replace with a urlparse-based path swap.
+"""
+
+from urllib.parse import urlparse, urlunparse
+
+
+def probe_url(nv_embed_url: str) -> str:
+    """Re-implementation of the probe URL builder from compat/server.py.
+    Kept tiny + isolated so it can be unit-tested without booting FastAPI."""
+    u = urlparse(nv_embed_url)
+    return urlunparse((u.scheme, u.netloc, "/health", "", "", ""))
+
+
+def test_probe_with_v1_embeddings_path():
+    assert probe_url("https://nv-embed:8041/v1/embeddings") == "https://nv-embed:8041/health"
+
+
+def test_probe_with_v1_embed_path():
+    """The bug case — gateway moved to /v1/embed and old replace() left URL untouched."""
+    assert probe_url("https://lambda-gateway.pentatonic.com/v1/embed") == (
+        "https://lambda-gateway.pentatonic.com/health"
+    )
+
+
+def test_probe_with_bare_host():
+    """No path at all — provider supplies path_default, but probe just wants /health."""
+    assert probe_url("https://lambda-gateway.pentatonic.com") == (
+        "https://lambda-gateway.pentatonic.com/health"
+    )
+
+
+def test_probe_strips_query_and_fragment():
+    assert probe_url("https://gw.example.com/v1/embeddings?token=x#frag") == (
+        "https://gw.example.com/health"
+    )
+
+
+def test_probe_preserves_port():
+    assert probe_url("http://nv-embed:8041/v1/embeddings") == "http://nv-embed:8041/health"

package/packages/memory-engine/tests/test_l6_module_load.py

@@ -0,0 +1,84 @@
+"""Regression test for the L6 _embed_client shadowing bug introduced in v0.8.0.
+
+When the EmbedClient refactor landed in 0.8.0, the new `def _embed_client()`
+factory function was added at module top, but the legacy module-level
+`_embed_client = httpx.Client(timeout=60)` binding (used by Ollama entity
+extraction) was left in place. Python's top-to-bottom evaluation rebound
+the name to the httpx.Client instance, so any subsequent call to
+`_embed_client()` raised `TypeError: 'Client' object is not callable`.
+
+This silently 500'd every L6 /index-batch and /search request from 0.8.0
+through 0.8.2 — the bug couldn't be caught by /health because the process
+itself stays up, only the request handlers fail.
+
+This is a static-source test (parses the file) rather than an import-time
+test because L6's heavy imports (pymilvus, spacy) aren't available in the
+unit-test venv. The check: scan the AST for any non-function rebinding of
+identifiers that are also defined as `def` in the same module. Catches
+this exact bug shape across any service that uses the EmbedClient factory
+pattern.
+"""
+
+from __future__ import annotations
+
+import ast
+from pathlib import Path
+
+import pytest
+
+SERVICES_DIR = Path(__file__).parent.parent / "engine" / "services"
+
+# Services that use the lazy EmbedClient factory pattern.
+SERVICES_WITH_EMBED_FACTORY = [
+    SERVICES_DIR / "l4" / "server.py",
+    SERVICES_DIR / "l5" / "l5-comms-layer.py",
+    SERVICES_DIR / "l6" / "l6-document-store.py",
+    SERVICES_DIR / "l2" / "l2-hybridrag-proxy.py",
+]
+
+
+def _module_level_defs_and_assigns(source: str) -> tuple[set[str], set[str]]:
+    """Return (function names, non-function-assigned names) at module level."""
+    tree = ast.parse(source)
+    funcs: set[str] = set()
+    assigns: set[str] = set()
+    for node in tree.body:
+        if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
+            funcs.add(node.name)
+        elif isinstance(node, ast.Assign):
+            for target in node.targets:
+                if isinstance(target, ast.Name):
+                    assigns.add(target.id)
+        elif isinstance(node, ast.AnnAssign) and isinstance(node.target, ast.Name):
+            assigns.add(node.target.id)
+    return funcs, assigns
+
+
+@pytest.mark.parametrize("service_file", SERVICES_WITH_EMBED_FACTORY, ids=lambda p: f"{p.parent.name}/{p.name}")
+def test_no_module_level_shadowing_of_factory_functions(service_file: Path):
+    """A module-level `def foo()` must not also have a module-level `foo = ...`
+    later in the file. That's exactly the shape that caused the v0.8.0 L6 bug."""
+    source = service_file.read_text()
+    funcs, assigns = _module_level_defs_and_assigns(source)
+    overlap = funcs & assigns
+    assert not overlap, (
+        f"{service_file.relative_to(SERVICES_DIR.parent.parent)} has module-level "
+        f"identifier(s) defined as both `def` and `name = ...`: {sorted(overlap)}. "
+        f"This causes silent name shadowing — the assignment wins and any call "
+        f"to {sorted(overlap)[0]}() raises TypeError at runtime."
+    )
+
+
+def test_l6_uses_renamed_ollama_http_not_embed_client():
+    """Belt-and-suspenders: explicitly assert L6's Ollama HTTP client is at
+    `_ollama_http`, not `_embed_client`. If someone reintroduces the original
+    binding by accident, this test catches it without depending on AST traversal."""
+    source = (SERVICES_DIR / "l6" / "l6-document-store.py").read_text()
+    assert "_embed_client = httpx.Client" not in source, (
+        "L6 reintroduced the legacy `_embed_client = httpx.Client(...)` binding "
+        "that shadowed the EmbedClient factory in v0.8.0. Rename to _ollama_http."
+    )
+    assert "_ollama_http = httpx.Client" in source, (
+        "L6 is missing the renamed Ollama HTTP client (`_ollama_http = httpx.Client(...)`). "
+        "The Ollama entity-extraction call path needs an httpx.Client somewhere."
+    )