WIP: initial vespa search support via cpr_data_access

Dockerfile

@@ -4,7 +4,7 @@ RUN mkdir /cpr-backend
 WORKDIR /cpr-backend
 
 RUN apt update && \
-    apt install -y postgresql-client curl \
+    apt install -y postgresql-client curl git \
     && rm -rf /var/lib/apt/lists/*
 
 # Install pip and poetry
@@ -18,6 +18,7 @@ COPY poetry.lock pyproject.toml ./
 RUN poetry export --with dev \
     | grep -v '\--hash' \
     | grep -v '^torch' \
+    | grep -v '^triton' \
     | grep -v '^nvidia' \
     | sed -e 's/ \\$//' \
     | sed -e 's/^[[:alpha:]]\+\[\([[:alpha:]]\+\[[[:alpha:]]\+\]\)\]/\1/' \
@@ -26,7 +27,7 @@ RUN poetry export --with dev \
 # e.g. we need to replace pydocstyle[pydocstyle[toml]] with pydocstyle[toml]
 
 # Install torch-cpu with pip
-RUN pip3 install --no-cache "torch==1.13.0+cpu" "torchvision==0.14.0+cpu" -f https://download.pytorch.org/whl/torch_stable.html
+RUN pip3 install --no-cache "torch==2.0.0+cpu" "torchvision==0.15.1+cpu" -f https://download.pytorch.org/whl/torch_stable.html
 
 # Install application requirements
 RUN pip3 install --no-cache -r requirements.txt

app/core/search.py

@@ -11,9 +11,9 @@
 from typing import Any, Mapping, Optional, Sequence, cast
 import string
 
+from cpr_data_access.embedding import Embedder
 from opensearchpy import OpenSearch
 from opensearchpy import JSONSerializer as jss
-from sentence_transformers import SentenceTransformer
 from sqlalchemy.orm import Session
 
 from app.api.api_v1.schemas.search import (
@@ -73,10 +73,10 @@
 
 _LOGGER = logging.getLogger(__name__)
 
-_ENCODER = SentenceTransformer(
-    model_name_or_path=OPENSEARCH_INDEX_ENCODER,
+_ENCODER = Embedder(
     cache_folder=os.environ.get("INDEX_ENCODER_CACHE_FOLDER", "/models"),
 )
+
 # Map a sort field type to the document key used by OpenSearch
 _SORT_FIELD_MAP: Mapping[SortField, str] = {
     SortField.DATE: "document_date",
@@ -410,7 +410,11 @@ def with_semantic_query(self, query_string: str, knn: bool):
 
         _LOGGER.info(f"Starting embeddings generation for '{query_string}'")
         start_generation = time.time_ns()
-        embedding = _ENCODER.encode(query_string, show_progress_bar=False)
+        embedding = _ENCODER.embed(
+            query_string,
+            normalize=False,
+            show_progress_bar=False,
+        )
         end_generation = time.time_ns()
         embeddings_generation_time = round((end_generation - start_generation) / 1e6)
         _LOGGER.info(