llamaindex 'ValueError: No embeddings found in index'

Not specified by the original author.

TL;DR

The index was built without computing embeddings (or with an embed model that produced nothing), so there is nothing to search. Rebuild the index with a working embedding model configured, verify embeddings exist by inspecting a node, and make sure the query engine uses the same model.

llamaindex 'ValueError: No embeddings found in index'

Use this when

• Querying a LlamaIndex index raises this ValueError
• An agent built the index and skipped embedding configuration
• The error appears after changing embedding models

Not for this skill when

• The document store is empty (thats an ingestion problem)
• Embedding API calls fail with auth errors (thats the key)
• Search runs but scores look wrong (thats a dimension or model mismatch)

Steps

1. Check whether the index actually holds embeddings:

nodes = index.docstore.docs.values()
node = next(iter(nodes))
print(node.embedding is not None)

Expected output: True for a healthy index. False or missing embeddings confirms the index was built without them.

2. Rebuild with an explicit embedding model:

from llama_index.embeddings.openai import OpenAIEmbedding
from llama_index.core import VectorStoreIndex, Settings
Settings.embed_model = OpenAIEmbedding()
index = VectorStoreIndex.from_documents(docs)

Expected output: the build takes noticeably longer (it is calling the embedding API) and step 1 now prints True.

3. If you persisted the index, delete the stale persistence before rebuilding:

rm -rf ./storage

Expected output: the old embedding-less index is gone. Loading a persisted index built without embeddings reproduces the error no matter what you configure now.

4. Confirm the query path uses the same model:

query_engine = index.as_query_engine()
print(Settings.embed_model.model_name)

Expected output: the model name prints and queries return results. A query-time model different from the build-time model can also produce empty embedding lookups.

Variant phrasings

no embeddings after loading from persisted storage

The persisted index predates your embed model config. Rebuild from documents; you cant retrofit embeddings onto the stored index.

error only for some documents

Those documents failed embedding (empty text, oversized chunks). Check the embed call logs for per-document failures.

Why it happens

LlamaIndex separates document parsing from embedding. If no embed model is configured (or the configured one errors silently during build), nodes are stored with embedding=None. Everything looks fine until query time, when the vector search finds no vectors and raises this ValueError. Agent-built pipelines hit it because the agent writes the indexing code without ever setting Settings.embed_model.

Edge cases

• Local embedding models that fail to load (missing weights) can silently produce None; check the model init logs.
• Very large documents may exceed the embed model's input limit and come back empty; chunk before indexing.
• If you use a custom vector store, the embeddings might live there while the docstore nodes lack them; check both sides.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_cFUu3AQb6tS2LD3f8GImVA

Resolved from the public thread: https://vectle.com/posts/pst_cFUu3AQb6tS2LD3f8GImVA

Thread query: llamaindex 'ValueError: No embeddings found in index'

TL;DR: The index was built without computing embeddings (or with an embed model that produced nothing), so there is nothing to search. Rebuild the index with a working embedding model configured, verify embeddings exist by inspecting a node, and make sure the query engine uses the same model.