Create app.py

app.py

@@ -0,0 +1,77 @@
+import json
+import numpy as np
+from fastapi import FastAPI
+from pydantic import BaseModel
+from huggingface_hub import hf_hub_download
+from transformers import AutoTokenizer, AutoModel
+import torch
+
+# -----------------------------
+# CONFIG
+# -----------------------------
+EMBEDDER_REPO = "ClergeF/MVT-embedder"
+MODEL_REPO = "ClergeF/impact-model"
+MODEL_FILE = "impact.json"
+
+# -----------------------------
+# SAFE EMBEDDER LOADER
+# -----------------------------
+def load_safe_embedder(repo_id: str):
+    print(f"Loading embedder from {repo_id} (safe mode)...")
+    tokenizer = AutoTokenizer.from_pretrained(repo_id)
+    model = AutoModel.from_pretrained(repo_id)
+
+    def embed_fn(texts):
+        tokens = tokenizer(
+            texts,
+            padding=True,
+            truncation=True,
+            max_length=256,
+            return_tensors="pt"
+        )
+        with torch.no_grad():
+            outputs = model(**tokens).last_hidden_state
+            embeddings = outputs.mean(dim=1).numpy()
+        return embeddings
+
+    return embed_fn
+
+# Load embedder
+embed_fn = load_safe_embedder(EMBEDDER_REPO)
+
+# -----------------------------
+# LOAD THE IMPACT MODEL
+# -----------------------------
+print("Loading impact model...")
+path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE)
+with open(path, "r") as f:
+    impact_model = json.load(f)
+
+# -----------------------------
+# HELPERS
+# -----------------------------
+def embed(text: str):
+    return embed_fn([text])[0]
+
+def linear_predict(model_json, vec):
+    coef = np.array(model_json["coef"])
+    intercept = np.array(model_json["intercept"])
+    return float(np.dot(coef, vec) + intercept)
+
+# -----------------------------
+# FASTAPI
+# -----------------------------
+app = FastAPI(title="Impact Model API")
+
+class Input(BaseModel):
+    text: str
+
+@app.post("/predict")
+def predict(payload: Input):
+    vec = embed(payload.text)
+    impact = linear_predict(impact_model, vec)
+
+    return {
+        "input": payload.text,
+        "impact_score": impact
+    }