run-llama/llama_index · error · ValueError

Cannot initialize from a vector store that does not store te

Error message

Cannot initialize from a vector store that does not store text.

What it means

MultiModalVectorStoreIndex.from_vector_store() reconstructs an index purely from an existing vector store, which requires the store to return the node text itself (stores_text=True). If the store only returns ids/embeddings (stores_text=False, e.g. many stores that rely on a separate docstore), there is no way to rebuild text and image nodes from it alone, so the constructor raises ValueError immediately.

Source

Thrown at llama-index-core/llama_index/core/indices/multi_modal/base.py:212

                retriever=self.as_retriever(**kwargs),
                multi_modal_llm=llm,
                **kwargs,
            )

        return super().as_chat_engine(chat_mode, llm, **kwargs)

    @classmethod
    def from_vector_store(
        cls,
        vector_store: BasePydanticVectorStore,
        embed_model: Optional[EmbedType] = None,
        # Image-related kwargs
        image_vector_store: Optional[BasePydanticVectorStore] = None,
        image_embed_model: EmbedType = "clip",
        **kwargs: Any,
    ) -> "MultiModalVectorStoreIndex":
        if not vector_store.stores_text:
            raise ValueError(
                "Cannot initialize from a vector store that does not store text."
            )

        storage_context = StorageContext.from_defaults(vector_store=vector_store)
        return cls(
            nodes=[],
            storage_context=storage_context,
            image_vector_store=image_vector_store,
            image_embed_model=image_embed_model,
            embed_model=(
                resolve_embed_model(
                    embed_model, callback_manager=kwargs.get("callback_manager")
                )
                if embed_model
                else Settings.embed_model
            ),
            **kwargs,
        )

View on GitHub (pinned to afd0fef371)

Solutions

1. Use a vector store that keeps text (stores_text=True), e.g. most local or metadata-retaining stores.
2. Or build the multimodal index with nodes: MultiModalVectorStoreIndex(nodes=..., vector_store=..., image_vector_store=...) so text is ingested rather than reconstructed.
3. Check the flag before calling: assert vector_store.stores_text, and pick the construction method accordingly.

Example fix

# before
index = MultiModalVectorStoreIndex.from_vector_store(vector_store=my_store)  # stores_text=False

# after
assert my_store.stores_text
index = MultiModalVectorStoreIndex.from_vector_store(vector_store=my_store)
# or ingest nodes directly:
# index = MultiModalVectorStoreIndex(nodes=nodes, vector_store=my_store)

assert vector_store.stores_text, (
    "from_vector_store requires a text-storing store; "
    "build with nodes=... instead"
)
index = MultiModalVectorStoreIndex.from_vector_store(vector_store=vector_store)

def usable_for_mm_from_vector_store(store) -> bool:
    return bool(getattr(store, "stores_text", False))

try:
    index = MultiModalVectorStoreIndex.from_vector_store(vector_store=store)
except ValueError as e:
    if "does not store text" in str(e):
        index = MultiModalVectorStoreIndex(nodes=nodes, vector_store=store)
    else:
        raise

When it happens

Trigger: Passing a stores_text=False vector store (e.g. some docstore-backed integrations) to MultiModalVectorStoreIndex.from_vector_store(); also note the image store path: image_vector_store has its own stores_text requirement in the surrounding code.

Common situations: Reusing a VectorStoreIndex setup with a docstore-dependent store for multimodal from_vector_store; switching a working single-modal pipeline to MultiModalVectorStoreIndex while keeping the same store; stores whose stores_text flag differs by version.

Related errors

• Unknown retriever mode: {retriever_mode}
• Unknown retriever mode: {retriever_mode}
• Vector store query result should return at least one of node
• llm must start with str 'local' or of type LLM or BaseLangua
• Token limit must be set and greater than 0.

AI-assisted analysis of run-llama/llama_index@afd0fef371 (2026-08-15). Data as JSON: /api/errors/3c6608dd5494a028. Report an issue: GitHub.