Tune GLiNER confidence floors (#114)

* Tune GLiNER confidence floors

* Fix Docker torch dependency install

diff --git a/detector/detector/gliner_layer.py b/detector/detector/gliner_layer.py
index 2b237560291bbe794b87865e868f4917d2c602fd..e9290c521a3e7ea30d2e421ce97ab76353d78b3a 100644 (file)
--- a/detector/detector/gliner_layer.py
+++ b/detector/detector/gliner_layer.py
@@ -26,11 +26,11 @@ def _floor(label: str, default: float) -> float:
 # overridable via env, e.g. DETECTOR_FLOOR_LOCATION=0.6). Role nouns are demoted
 # by the suppressor labels (below), not by this floor.
 PER_LABEL_FLOOR = {
-    "person": _floor("person", 0.85),
-    "location": _floor("location", 0.50),
+    "person": _floor("person", 0.95),
+    "location": _floor("location", 0.80),
     # A dedicated "address" label recovers full street addresses that a bare
     # "location" reading misses; emitted as LOCATION (see _LABEL_TO_TYPE).
-    "address": _floor("address", 0.70),
+    "address": _floor("address", 0.80),
 }
 # Labels the request `score_threshold` may raise (high-volume, deployment-tunable).
 _TUNABLE = {"person", "location", "address"}

diff --git a/docker/Dockerfile b/docker/Dockerfile
index bc14936cc9096cb79532f76711ae4b6ec4e1e4cc..7083cb97f7f8f235db7d7f8e9cb60af46cb5b108 100644 (file)
--- a/docker/Dockerfile
+++ b/docker/Dockerfile
@@ -37,7 +37,8 @@ FROM python:3.11-slim AS detector
 # pin it for every arch. (PyPI's DEFAULT index now ships CUDA builds for
 # linux/arm64 too — e.g. torch 2.x+cuXXX — which would drag ~6 GB of unused
 # CUDA libraries into the image; the explicit CPU index avoids that.)
-RUN pip install --no-cache-dir torch --index-url https://download.pytorch.org/whl/cpu
+RUN pip install --no-cache-dir typing-extensions \
+    && pip install --no-cache-dir torch --index-url https://download.pytorch.org/whl/cpu
 
 # Install the detector package (pulls fastapi/uvicorn/gliner/stdnum/...).
 COPY detector/pyproject.toml /srv/detector/