Set up devis-matcher encoder API

README.md

@@ -1,15 +1,31 @@
 ---
-title: Test
-emoji: 🐢
-colorFrom: gray
-colorTo: indigo
+title: Devis Matcher Encoder
+emoji: 🪛
+colorFrom: indigo
+colorTo: pink
 sdk: gradio
-sdk_version: 6.14.0
-python_version: '3.13'
+sdk_version: 4.44.0
 app_file: app.py
 pinned: false
 license: mit
-short_description: test
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Devis Matcher Encoder
+
+HF Space exposing `Pradhap/devis-matcher` (fine-tuned camembert-large, ~440 MB)
+as a sentence-embedding API for French construction document matching.
+
+Used by the ConstructCRM backend (FastAPI on Render) to keep heavy model
+weights off the 512 MB worker.
+
+## API
+
+```
+POST /api/encode/
+Content-Type: application/json
+Body:    {"data": [["text 1", "text 2"]]}
+Returns: {"data": [[[float, ...], [float, ...]]], ...}
+```
+
+Send as many strings as you can in one call — the model encodes batches
+significantly faster than one-at-a-time.

app.py

@@ -1,7 +1,59 @@
+"""HF Space exposing Pradhap/devis-matcher as a sentence-embedding API.
+
+The fine-tuned camembert-large model (~440 MB) is loaded once at Space startup
+into the container's 16 GB RAM, then used to encode batches of strings on
+demand. The ConstructCRM backend (Render free tier, 512 MB) calls this Space
+instead of loading the model locally — which would OOM the worker.
+
+API surface (auto-generated by Gradio):
+
+  POST /api/encode/
+       body: {"data": [["text 1", "text 2"]]}
+    returns: {"data": [[[float, ...], [float, ...]]], ...}
+
+For batch efficiency, send as many texts as you can in a single call.
+"""
 import gradio as gr
+from sentence_transformers import SentenceTransformer
+
+MODEL_NAME = "Pradhap/devis-matcher"
+
+print(f"[Space] Loading {MODEL_NAME} (cold start may take a minute)…")
+model = SentenceTransformer(MODEL_NAME)
+print(f"[Space] Loaded. Embedding dim: {model.get_sentence_embedding_dimension()}")
+
+
+def encode(texts):
+    """Encode a list of strings into sentence embeddings."""
+    if texts is None:
+        return []
+    if isinstance(texts, str):
+        texts = [texts]
+    if not isinstance(texts, list):
+        return {"error": f"expected list[str], got {type(texts).__name__}"}
+    texts = [t for t in texts if isinstance(t, str)]
+    if not texts:
+        return []
+    embs = model.encode(texts, convert_to_numpy=True, normalize_embeddings=False)
+    return embs.tolist()
+
 
-def greet(name):
-    return "Hello " + name + "!!"
+demo = gr.Interface(
+    fn=encode,
+    inputs=gr.JSON(label="Texts (JSON array of strings)"),
+    outputs=gr.JSON(label="Embeddings (2D float array)"),
+    title="Devis Matcher Encoder",
+    description=(
+        f"Sentence embeddings via **{MODEL_NAME}** "
+        "(camembert-large fine-tuned on French construction documents). "
+        "POST a list of strings, receive a 2D array of embedding vectors."
+    ),
+    examples=[
+        [["PEINTURE SUR MURS", "Carrelage faïence salle de bain", "Plomberie cuisine"]],
+    ],
+    api_name="encode",
+    allow_flagging="never",
+)
 
-demo = gr.Interface(fn=greet, inputs="text", outputs="text")
-demo.launch()
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,3 @@
+sentence-transformers>=2.7.0
+gradio>=4.44.0
+torch>=2.0.0