Give each Chroma store its own Settings; a shared one crossed the bundles

src/retrievers/csv_chroma.py

@@ -21,7 +21,32 @@
 from nltk.tokenize import word_tokenize
 from pydantic import ConfigDict
 
-chroma_settings = chromadb.config.Settings(anonymized_telemetry=False)
+
+def chroma_settings() -> chromadb.config.Settings:
+    """A *fresh* Settings object for every Chroma store.
+
+    Not a shared constant, which is what this was. `langchain_chroma` mutates
+    whatever it is handed:
+
+        client_settings.persist_directory = (
+            persist_directory or client_settings.persist_directory
+        )
+
+    so one shared instance means the last store constructed rewrites the
+    persist directory for every earlier one, and chromadb hands back a client
+    keyed on those settings. With only the reactome bundle installed nothing
+    showed, because every store pointed at the same tree. Installing the user
+    guide bundle gave them different trees, and a reactome question started
+    looking for its collection inside the user guide directory:
+
+        PermissionError: [Errno 13] Permission denied:
+        '/app/embeddings/.../userguide/Release95/sections/9c574827-...'
+
+    -- Chroma trying to create a segment folder for a collection that is not
+    there, in a bundle it should never have opened.
+    """
+    return chromadb.config.Settings(anonymized_telemetry=False)
+
 
 multi_query_prompt = PromptTemplate(
     input_variables=["question"],
@@ -274,7 +299,7 @@ def from_subdirectory(
             vectordb = Chroma(
                 persist_directory=str(embeddings_directory / subdirectory),
                 embedding_function=embedding,
-                client_settings=chroma_settings,
+                client_settings=chroma_settings(),
             )
             vector_retriever = vectordb.as_retriever(
                 search_kwargs={"k": RESULTS_PER_RETRIEVER * VECTOR_OVERFETCH}

src/retrievers/userguide/retriever.py

@@ -32,6 +32,6 @@ def create_userguide_retriever(
     vectordb = Chroma(
         persist_directory=str(chroma_path),
         embedding_function=embedding,
-        client_settings=chroma_settings,
+        client_settings=chroma_settings(),
     )
     return vectordb.as_retriever(search_kwargs={"k": k})

tests/retrievers/test_chroma_settings.py

@@ -0,0 +1,48 @@
+"""Each Chroma store needs its own Settings object.
+
+`langchain_chroma` mutates whatever settings it is handed:
+
+    client_settings.persist_directory = (
+        persist_directory or client_settings.persist_directory
+    )
+
+so a shared instance means the last store constructed rewrites the persist
+directory for every earlier one, and chromadb returns a client keyed on those
+settings.
+
+With only the reactome bundle installed this was invisible -- every store
+pointed into the same tree. Installing the user guide bundle gave them
+different trees, and a *reactome* question went looking for its collection
+inside the *user guide* directory:
+
+    PermissionError: [Errno 13] Permission denied:
+    '/app/embeddings/.../userguide/Release95/sections/9c574827-...'
+
+Found on beta, in the deployed container, by asking "What does CDK5
+phosphorylate in Alzheimer disease?" after a user guide question.
+"""
+
+from retrievers.csv_chroma import chroma_settings
+
+
+def test_each_call_returns_a_distinct_object() -> None:
+    first, second = chroma_settings(), chroma_settings()
+    assert first is not second, (
+        "a shared Settings object is mutated by langchain_chroma, so two stores "
+        "would end up pointing at one directory"
+    )
+
+
+def test_mutating_one_does_not_reach_another() -> None:
+    """This is the exact mutation langchain_chroma performs."""
+    first = chroma_settings()
+    first.persist_directory = "/bundles/userguide/sections"
+
+    second = chroma_settings()
+    assert second.persist_directory != "/bundles/userguide/sections"
+
+
+def test_telemetry_stays_off() -> None:
+    # The reason this object existed in the first place; do not lose it while
+    # fixing the sharing.
+    assert chroma_settings().anonymized_telemetry is False