amfs-http-server 0.3.4__tar.gz → 0.3.6__tar.gz

{amfs_http_server-0.3.4 → amfs_http_server-0.3.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: amfs-http-server
-Version: 0.3.4
+Version: 0.3.6
 Summary: AMFS HTTP/REST API server with SSE support
 License-Expression: Apache-2.0
 Requires-Python: >=3.11

{amfs_http_server-0.3.4 → amfs_http_server-0.3.6}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "amfs-http-server"
-version = "0.3.4"
+version = "0.3.6"
 description = "AMFS HTTP/REST API server with SSE support"
 requires-python = ">=3.11"
 license = "Apache-2.0"

{amfs_http_server-0.3.4 → amfs_http_server-0.3.6}/src/amfs_http/pro_provider.py

@@ -14,7 +14,13 @@ Everything here is soft/optional and fail-open:
 
 Env:
 - AMFS_PROXY_PRO_RETRIEVAL   "false" to force the OSS search path (default true)
-- AMFS_PROXY_PRO_RERANK      "false" to skip cross-encoder rerank (default true)
+- AMFS_PROXY_PRO_RERANK      "false" to skip reranking entirely (default true)
+- AMFS_LLM_RERANK            "true" to use the LLM listwise reranker (Pro tier) —
+                             the lever that reaches >90% adversarial Hit@1. Costs
+                             one LLM call per retrieval, so it is opt-in. Needs
+                             AMFS_LLM_RERANK_BASE_URL/API_KEY (e.g. OpenRouter).
+                             When unset/unavailable, falls back to the offline
+                             cross-encoder.
 
 NOTE(integration-test): the Pro retrieval path requires the amfs-internal packages
 and a populated adapter; it is exercised in the amfs-internal retrieval eval, not
@@ -35,6 +41,34 @@ PRO_RERANK = os.environ.get("AMFS_PROXY_PRO_RERANK", "true").lower() == "true"
 _cache: dict[str, Any] = {}
 
 
+def _build_reranker() -> Any | None:
+    """Build the strongest available reranker.
+
+    Prefers the LLM listwise reranker (Pro tier) when ``AMFS_LLM_RERANK=true`` and
+    an OpenAI-compatible client/key are configured — it resolves confusable facts
+    under adversarial paraphrase and is the lever for top-1 accuracy, at the cost
+    of one LLM call per retrieval. Otherwise falls back to the offline
+    ``CrossEncoderReranker`` (single-digit ms, zero egress). Returns ``None`` only
+    if neither is importable.
+    """
+    try:
+        from amfs_retrieval import LLMReranker
+
+        llm = LLMReranker()  # reads AMFS_LLM_RERANK* env
+        if llm.available:
+            logger.info("pro_provider: LLM listwise reranker enabled for proxy")
+            return llm
+    except Exception:  # noqa: BLE001 - LLM rerank optional
+        logger.debug("pro_provider: LLM reranker unavailable", exc_info=True)
+    try:
+        from amfs_retrieval import CrossEncoderReranker
+
+        return CrossEncoderReranker()
+    except Exception:  # noqa: BLE001 - rerank optional
+        logger.debug("pro_provider: reranker unavailable", exc_info=True)
+        return None
+
+
 def build_pro_retriever(adapter: Any) -> Any | None:
     """Build a Pro MultiStrategyRetriever over ``adapter``, or None if unavailable.
 
@@ -48,14 +82,7 @@ def build_pro_retriever(adapter: Any) -> Any | None:
     try:
         from amfs_retrieval import MultiStrategyRetriever, create_pro_embedder
 
-        reranker = None
-        if PRO_RERANK:
-            try:
-                from amfs_retrieval import CrossEncoderReranker
-
-                reranker = CrossEncoderReranker()
-            except Exception:  # noqa: BLE001 - rerank optional
-                logger.debug("pro_provider: reranker unavailable", exc_info=True)
+        reranker = _build_reranker() if PRO_RERANK else None
 
         retriever = MultiStrategyRetriever(
             adapter, embedder=create_pro_embedder(), reranker=reranker

{amfs_http_server-0.3.4 → amfs_http_server-0.3.6}/src/amfs_http/server.py

@@ -66,6 +66,27 @@ from amfs_http.sse import SSEManager
 
 logger = logging.getLogger(__name__)
 
+
+def _server_version() -> str:
+    """Return the deployed amfs-http-server package version.
+
+    Sourced from installed package metadata so it changes on every release —
+    unlike the per-entry ``amfs_version`` schema tag. This is the value to use
+    when telling deploys/revisions apart.
+    """
+    try:
+        from importlib.metadata import version
+
+        return version("amfs-http-server")
+    except Exception:
+        return "unknown"
+
+
+# Set by the deploy pipeline (e.g. the git SHA) so a running revision is
+# identifiable even between version bumps. Empty when not provided.
+_BUILD_SHA = os.environ.get("AMFS_BUILD_SHA", "")
+_SCHEMA_VERSION = MemoryEntry.model_fields["amfs_version"].default
+
 # ── Async adapter (hot-path, non-blocking) ──────────────────────────
 _async_adapter = None  # AsyncPostgresAdapter | None, set in lifespan
 
@@ -94,7 +115,7 @@ async def _lifespan(application: FastAPI):  # noqa: ARG001
 app = FastAPI(
     title="AMFS HTTP API",
     description="Agent Memory File System — REST API with SSE support",
-    version="0.1.0",
+    version=_server_version(),
     lifespan=_lifespan,
 )
 
@@ -316,14 +337,25 @@ def _ensure_agent_owner(
 # ──────────────────────────────────────────────────────────────────────
 
 
+def _health_payload() -> dict[str, str]:
+    payload = {
+        "status": "ok",
+        "version": _server_version(),      # deploy/package version — changes per release
+        "schema_version": _SCHEMA_VERSION,  # per-entry MemoryEntry schema tag
+    }
+    if _BUILD_SHA:
+        payload["build"] = _BUILD_SHA
+    return payload
+
+
 @app.get("/health")
 async def health() -> dict[str, str]:
-    return {"status": "ok"}
+    return _health_payload()
 
 
 @app.get("/api/v1/health")
 async def health_v1() -> dict[str, str]:
-    return {"status": "ok"}
+    return _health_payload()
 
 
 @app.get("/api/v1/auth/whoami")