Decouple embedding function from collection

doc/VectorCode-cli.txt

@@ -81,7 +81,7 @@ If you need to install multiple dependency group (for |VectorCode-cli-lsp| or
 |VectorCode-cli-mcp|), you can use the following syntax:
 
 >bash
-    uv tool install vectorcode[lsp,mcp]
+    uv tool install 'vectorcode[lsp,mcp]'
 <
 
 
@@ -136,7 +136,7 @@ LEGACY ENVIRONMENTS ~
 
 If your environment doesn’t support `numpy` version 2.0+, the default,
 unconstrained numpy may not work for you. In this case, you can try installing
-the package by `uv tool install vectorcode[legacy]`, which enforces numpy
+the package by `uv tool install 'vectorcode[legacy]'`, which enforces numpy
 `v1.x`. If this doesn’t help, please open an issue with your OS, CPU
 architecture, python version and the vectorcode virtual environment (`uv tool
 run --from=vectorcode python -m ensurepip && uv tool run --from=vectorcode
@@ -626,9 +626,9 @@ following options in the JSON config file:
 For Intel users, sentence transformer <https://www.sbert.net/index.html>
 supports OpenVINO
 <https://www.intel.com/content/www/us/en/developer/tools/openvino-toolkit/overview.html>
-backend for supported GPU. Run `uv install vectorcode[intel]` which will bundle
-the relevant libraries when you install VectorCode. After that, you will need
-to configure `SentenceTransformer` to use `openvino` backend. In your
+backend for supported GPU. Run `uv install 'vectorcode[intel]'` which will
+bundle the relevant libraries when you install VectorCode. After that, you will
+need to configure `SentenceTransformer` to use `openvino` backend. In your
 `config.json`, set `backend` key in `embedding_params` to `"openvino"`
 
 >json
@@ -760,11 +760,11 @@ The experimental language server can be installed via the `lsp` dependency
 group:
 
 >bash
-    pipx install vectorcode[lsp]
+    pipx install 'vectorcode[lsp]'
 
     ## or if you have an existing `vectorcode` install:
 
-    pipx inject vectorcode vectorcode[lsp] --force
+    pipx inject vectorcode 'vectorcode[lsp]' --force
 <
 
 The LSP request for the `workspace/executeCommand` is defined as follows:

src/vectorcode/cli_utils.py

@@ -114,6 +114,9 @@ class Config:
     files_action: Optional[FilesAction] = None
     rm_paths: list[str] = field(default_factory=list)
 
+    def __hash__(self) -> int:
+        return hash(self.__repr__())
+
     @classmethod
     async def import_from(cls, config_dict: dict[str, Any]) -> "Config":
         """

src/vectorcode/common.py

@@ -8,6 +8,7 @@
 import sys
 from asyncio.subprocess import Process
 from dataclasses import dataclass
+from functools import cache
 from typing import Any, AsyncGenerator, Optional
 from urllib.parse import urlparse
 
@@ -127,11 +128,15 @@ def get_collection_name(full_path: str) -> str:
     return collection_id
 
 
-def get_embedding_function(configs: Config) -> chromadb.EmbeddingFunction | None:
+@cache
+def get_embedding_function(configs: Config) -> chromadb.EmbeddingFunction:
     try:
-        return getattr(embedding_functions, configs.embedding_function)(
+        ef = getattr(embedding_functions, configs.embedding_function)(
             **configs.embedding_params
         )
+        if ef is None:  # pragma: nocover
+            raise AttributeError()
+        return ef
     except AttributeError:
         logger.warning(
             f"Failed to use {configs.embedding_function}. Falling back to Sentence Transformer.",
@@ -161,7 +166,6 @@ async def get_collection(
     full_path = str(expand_path(str(configs.project_root), absolute=True))
     if __COLLECTION_CACHE.get(full_path) is None:
         collection_name = get_collection_name(full_path)
-        embedding_function = get_embedding_function(configs)
 
         collection_meta: dict[str, str | int] = {
             "path": full_path,
@@ -183,14 +187,11 @@ async def get_collection(
             f"Getting/Creating collection with the following metadata: {collection_meta}"
         )
         if not make_if_missing:
-            __COLLECTION_CACHE[full_path] = await client.get_collection(
-                collection_name, embedding_function
-            )
+            __COLLECTION_CACHE[full_path] = await client.get_collection(collection_name)
         else:
             collection = await client.get_or_create_collection(
                 collection_name,
                 metadata=collection_meta,
-                embedding_function=embedding_function,
             )
             if (
                 not collection.metadata.get("hostname") == socket.gethostname()

src/vectorcode/subcommands/query/__init__.py

@@ -19,6 +19,7 @@
 from vectorcode.common import (
     ClientManager,
     get_collection,
+    get_embedding_function,
     verify_ef,
 )
 from vectorcode.subcommands.query.reranker import (
@@ -67,7 +68,7 @@ async def get_query_result_files(
                 )
                 logger.info(f"Querying {num_query} chunks for reranking.")
         results = await collection.query(
-            query_texts=query_chunks,
+            query_embeddings=get_embedding_function(configs)(query_chunks),
             n_results=num_query,
             include=[
                 IncludeEnum.metadatas,

src/vectorcode/subcommands/vectorise.py

@@ -28,6 +28,7 @@
 from vectorcode.common import (
     ClientManager,
     get_collection,
+    get_embedding_function,
     list_collection_files,
     verify_ef,
 )
@@ -92,6 +93,7 @@ async def chunked_add(
     max_batch_size: int,
     semaphore: asyncio.Semaphore,
 ):
+    embedding_function = get_embedding_function(configs)
     full_path_str = str(expand_path(str(file_path), True))
     orig_sha256 = None
     new_sha256 = hash_file(full_path_str)
@@ -147,6 +149,9 @@ async def chunked_add(
                     await collection.add(
                         ids=[get_uuid() for _ in inserted_chunks],
                         documents=[str(i) for i in inserted_chunks],
+                        embeddings=embedding_function(
+                            list(str(c) for c in inserted_chunks)
+                        ),
                         metadatas=metas,
                     )
     except (UnicodeDecodeError, UnicodeError):  # pragma: nocover

tests/subcommands/query/test_query.py

@@ -64,8 +64,14 @@ def mock_config():
 
 @pytest.mark.asyncio
 async def test_get_query_result_files(mock_collection, mock_config):
-    # Mock the reranker
-    with patch("vectorcode.subcommands.query.get_reranker") as mock_get_reranker:
+    mock_embedding_function = MagicMock()
+    with (
+        patch("vectorcode.subcommands.query.get_reranker") as mock_get_reranker,
+        patch(
+            "vectorcode.subcommands.query.get_embedding_function",
+            return_value=mock_embedding_function,
+        ),
+    ):
         mock_reranker_instance = MagicMock()
         mock_reranker_instance.rerank = AsyncMock(
             return_value=[
@@ -82,9 +88,7 @@ async def test_get_query_result_files(mock_collection, mock_config):
         # Check that query was called with the right parameters
         mock_collection.query.assert_called_once()
         args, kwargs = mock_collection.query.call_args
-        assert kwargs["query_texts"] == [
-            "test query"
-        ]  # Assuming chunking produces this
+        mock_embedding_function.assert_called_once_with(["test query"])
         assert kwargs["n_results"] == 6  # n_result(3) * query_multiplier(2)
         assert IncludeEnum.metadatas in kwargs["include"]
         assert IncludeEnum.distances in kwargs["include"]
@@ -285,10 +289,14 @@ async def test_get_query_result_files_chunking(mock_collection, mock_config):
     mock_config.query = [
         "this is a longer query that should be chunked into multiple parts"
     ]
-
+    mock_embedding_function = MagicMock()
     with (
         patch("vectorcode.subcommands.query.StringChunker") as MockChunker,
         patch("vectorcode.subcommands.query.reranker.NaiveReranker") as MockReranker,
+        patch(
+            "vectorcode.subcommands.query.get_embedding_function",
+            return_value=mock_embedding_function,
+        ),
     ):
         # Set up MockChunker to chunk the query
         mock_chunker_instance = MagicMock()
@@ -309,7 +317,7 @@ async def test_get_query_result_files_chunking(mock_collection, mock_config):
         # Check query was called with chunked query
         mock_collection.query.assert_called_once()
         _, kwargs = mock_collection.query.call_args
-        assert kwargs["query_texts"] == ["chunk1", "chunk2", "chunk3"]
+        mock_embedding_function.assert_called_once_with(["chunk1", "chunk2", "chunk3"])
 
         # Check the result
         assert result == ["file1.py", "file2.py"]
@@ -319,10 +327,14 @@ async def test_get_query_result_files_chunking(mock_collection, mock_config):
 async def test_get_query_result_files_multiple_queries(mock_collection, mock_config):
     # Set multiple query terms
     mock_config.query = ["term1", "term2", "term3"]
-
+    mock_embedding_function = MagicMock()
     with (
         patch("vectorcode.subcommands.query.StringChunker") as MockChunker,
         patch("vectorcode.subcommands.query.reranker.NaiveReranker") as MockReranker,
+        patch(
+            "vectorcode.subcommands.query.get_embedding_function",
+            return_value=mock_embedding_function,
+        ),
     ):
         # Set up MockChunker to return the query terms as is
         mock_chunker_instance = MagicMock()
@@ -342,7 +354,7 @@ async def test_get_query_result_files_multiple_queries(mock_collection, mock_con
         # Check query was called with all query terms
         mock_collection.query.assert_called_once()
         _, kwargs = mock_collection.query.call_args
-        assert set(kwargs["query_texts"]) == set(["term1", "term2", "term3"])
+        mock_embedding_function.assert_called_once_with(["term1", "term2", "term3"])
 
         # Check the result
         assert result == ["file1.py", "file2.py"]

tests/subcommands/test_vectorise.py

@@ -6,6 +6,7 @@
 from contextlib import ExitStack
 from unittest.mock import AsyncMock, MagicMock, mock_open, patch
 
+import numpy
 import pytest
 from chromadb.api.models.AsyncCollection import AsyncCollection
 from tree_sitter import Point
@@ -511,6 +512,7 @@ def is_file_side_effect(path):
         else:
             return True
 
+    mock_embedding_function = MagicMock(return_value=numpy.random.random((100,)))
     with (
         patch("os.path.isfile", side_effect=is_file_side_effect),
         patch(
@@ -522,6 +524,10 @@ def is_file_side_effect(path):
             "vectorcode.subcommands.vectorise.get_collection",
             return_value=mock_collection,
         ),
+        patch(
+            "vectorcode.subcommands.vectorise.get_embedding_function",
+            return_value=mock_embedding_function,
+        ),
         patch("vectorcode.subcommands.vectorise.verify_ef", return_value=True),
         patch(
             "vectorcode.subcommands.vectorise.expand_globs",