Move to remote embedder API

.dockerignore

@@ -25,6 +25,7 @@ coverage.xml
 *.md
 !README.md
 ingestion_pipeline
+embedder
 exefiles
 diagram
 diagram.svg

.gitignore

@@ -3,4 +3,5 @@ diagram
 
 # For ingestion pipeline
 ingestion_pipeline
+embedder 
 exefiles

Dockerfile

@@ -23,25 +23,12 @@ COPY . .
 # Install Python dependencies
 RUN pip install --upgrade pip && pip install --no-cache-dir -r requirements.txt
 
-# Hugging Face cache directories
+# Optional: general HF cache directory (kept for other models like BLIP)
 ENV HF_HOME="/home/user/.cache/huggingface"
-ENV SENTENCE_TRANSFORMERS_HOME="/home/user/.cache/huggingface/sentence-transformers"
-ENV MEDGEMMA_HOME="/home/user/.cache/huggingface/sentence-transformers"
 
-# Create cache directories and set permissions
-RUN mkdir -p /app/model_cache /home/user/.cache/huggingface/sentence-transformers && \
-    chown -R user:user /app/model_cache /home/user/.cache/huggingface
-
-# Control preloading flags
-ENV PRELOAD_TRANSLATORS="0"
-ENV EMBEDDING_HALF="0"
-
-# Preload embedding model and warmup
-RUN test -f /app/dw_model.py && python /app/dw_model.py || true
-RUN test -f /app/warmup.py && python /app/warmup.py || true
-
-# Ensure ownership stays correct
-RUN chown -R user:user /app/model_cache
+# Ensure cache directory ownership
+RUN mkdir -p /home/user/.cache/huggingface && \
+    chown -R user:user /home/user/.cache/huggingface
 
 # Expose port for HF Spaces
 ENV PORT=7860

dw_model.py

@@ -1,30 +0,0 @@
-# dw_model.py
-### --- A. transformer and embedder ---
-import os
-import shutil
-from huggingface_hub import snapshot_download
-
-# Set up paths
-MODEL_REPO = "sentence-transformers/all-MiniLM-L6-v2"
-MODEL_CACHE_DIR = "/app/model_cache"
-HF_CACHE_DIR = os.getenv("HF_HOME", "/home/user/.cache/huggingface")
-
-print("⏳ Downloading the SentenceTransformer model...")
-# Download directly into /app/model_cache to avoid duplicating files from HF cache
-model_path = snapshot_download(
-    repo_id=MODEL_REPO,
-    cache_dir=HF_CACHE_DIR,              # Store HF cache in user cache dir
-    local_dir=MODEL_CACHE_DIR,           # Place usable model here
-    local_dir_use_symlinks=False         # Copy files into local_dir (no symlinks)
-)
-
-print("Model path: ", model_path)
-if not os.path.exists(MODEL_CACHE_DIR):
-    os.makedirs(MODEL_CACHE_DIR)
-
-# Verify structure after moving
-print("\n📂 LLM Model Structure (Build Level):")
-for root, dirs, files in os.walk(MODEL_CACHE_DIR):
-    print(f"📁 {root}/")
-    for file in files:
-        print(f"  📄 {file}")

requirements.txt

@@ -8,7 +8,6 @@ PyMuPDF==1.24.10
 pillow==10.4.0
 transformers==4.44.2
 torch==2.2.2
-sentence-transformers==3.1.1
 sumy==0.11.0
 numpy==1.26.4
 reportlab==4.0.9

utils/rag/embeddings.py

@@ -2,32 +2,36 @@
 import os
 from typing import List
 import numpy as np
+import httpx
 from ..logger import get_logger
 
-try:
-    from sentence_transformers import SentenceTransformer
-except Exception:
-    SentenceTransformer = None
-
 
 logger = get_logger("EMBED", __name__)
 
 
 class EmbeddingClient:
-    def __init__(self, model_name: str = "sentence-transformers/all-MiniLM-L6-v2"):
+    def __init__(self, model_name: str = "sentence-transformers/all-MiniLM-L6-v2", api_url: str | None = None):
         self.model_name = model_name
-        self.model = None
-
-    def _lazy(self):
-        if self.model is None and SentenceTransformer is not None:
-            logger.info(f"Loading embedding model: {self.model_name}")
-            self.model = SentenceTransformer(self.model_name)
+        self.api_url = api_url or os.getenv("EMBEDDER_URL")
 
     def embed(self, texts: List[str]) -> List[list]:
-        self._lazy()
-        if self.model is None:
-            # Fallback: extremely naive hashing -> NOT for production, but keeps code running without deps
-            logger.warning("SentenceTransformer unavailable; using random fallback embeddings.")
+        if not texts:
+            return []
+
+        if not self.api_url:
+            logger.warning("EMBEDDER_URL not set; using random fallback embeddings.")
             return [list(np.random.default_rng(hash(t) % (2**32)).normal(size=384).astype("float32")) for t in texts]
-        vecs = self.model.encode(texts, show_progress_bar=False, normalize_embeddings=True)
-        return [v.tolist() for v in vecs]
+
+        url = self.api_url.rstrip("/") + "/embed"
+        try:
+            with httpx.Client(timeout=30.0) as client:
+                resp = client.post(url, json={"texts": texts})
+                resp.raise_for_status()
+                data = resp.json()
+                vectors = data.get("vectors")
+                if not isinstance(vectors, list):
+                    raise ValueError("Invalid response: 'vectors' field missing or not a list")
+                return vectors
+        except Exception as e:
+            logger.error(f"Embedding API call failed: {e}; falling back to random embeddings.")
+            return [list(np.random.default_rng(hash(t) % (2**32)).normal(size=384).astype("float32")) for t in texts]

warmup.py

@@ -1,17 +0,0 @@
-from sentence_transformers import SentenceTransformer
-import torch
-import os
-
-print("🚀 Warming up model...")
-embedding_model = SentenceTransformer("/app/model_cache", device="cpu")
-
-# Some CPU backends on HF Spaces fail on .half(); make it configurable
-USE_HALF = os.getenv("EMBEDDING_HALF", "1") == "1"
-try:
-    if USE_HALF and torch.cuda.is_available():
-        embedding_model = embedding_model.half()
-except Exception as e:
-    print(f"⚠️ Skipping half precision due to: {e}")
-
-embedding_model.to(torch.device("cpu"))
-print("✅ Model warm-up complete!")