fix(llm): failed to remove vectors when updating vid embedding

[MrJs133]

Bug Report

After deleting vertices, running update_vid_embedding does not successfully remove the corresponding vectors.

Initial State

• 13 vertices
• 13 vid embeddings

After Clearing Graph Data

• 0 vertices

After update_vid_embedding

• Output shows: “Removed 13 vectors”
  

However, calling get_vector_index_info still shows 13 vid embeddings.

---

Problem Location

Function get_vector_index_info:

def get_vector_index_info():
    chunk_vector_index = VectorIndex.from_index_file(
        os.path.join(resource_path, huge_settings.graph_name, "chunks")
    )
    graph_vid_vector_index = VectorIndex.from_index_file(
        os.path.join(resource_path, huge_settings.graph_name, "graph_vids")
    )
    return json.dumps({
        "embed_dim": chunk_vector_index.index.d,
        "vector_info": {
            "chunk_vector_num": chunk_vector_index.index.ntotal,
            "graph_vid_vector_num": graph_vid_vector_index.index.ntotal,
            "graph_properties_vector_num": len(chunk_vector_index.properties)
        }
    }, ensure_ascii=False, indent=2)

This logic is correct. It reads graph_vid_vector_num from graph_vid_vector_index.index.ntotal.

update_vid_embedding Code Analysis

past_vids = self.vid_index.properties
present_vids = context["vertices"]
removed_vids = set(past_vids) - set(present_vids)
removed_num = self.vid_index.remove(removed_vids)
added_vids = list(set(present_vids) - set(past_vids))

This correctly identifies vectors to be removed and added.
self.vid_index.remove() implementation:

def remove(self, props: Union[Set[Any], List[Any]]) -> int:
    if isinstance(props, list):
        props = set(props)
    indices = []
    remove_num = 0

    for i, p in enumerate(self.properties):
        if p in props:
            indices.append(i)
            remove_num += 1
    self.index.remove_ids(np.array(indices))
    self.properties = [p for i, p in enumerate(self.properties) if i not in indices]
    return remove_num

This also seems correct.

---

Debug Output

log.debug("before %s", self.vid_index.index.ntotal)
removed_num = self.vid_index.remove(removed_vids)
log.debug("after %s", self.vid_index.index.ntotal)

Output:

[05/20/25 13:51:20] DEBUG before 13
[05/20/25 13:51:20] DEBUG after 0

→ This confirms that in-memory deletion is successful.

However, re-running update_vid_embedding again shows:

[05/20/25 13:53:23] DEBUG before 13
[05/20/25 13:53:23] DEBUG after 0

→ Confirms that the vector index file still contains 13 vectors (i.e., deletion was not persisted).

And this is verified by loading the index from file via:

self.index_dir = os.path.join(resource_path, huge_settings.graph_name, "graph_vids")
self.vid_index = VectorIndex.from_index_file(self.index_dir)

log.debug("after %s", self.vid_index.index.ntotal)

Note: The result of this deletion was not saved.

Root Cause

In the full BuildSemanticIndex.run() implementation:

removed_vids = set(past_vids) - set(present_vids)
removed_num = self.vid_index.remove(removed_vids)
added_vids = list(set(present_vids) - set(past_vids))

if added_vids:
    ...
    self.vid_index.add(...)
    self.vid_index.to_index_file(self.index_dir)
else:
    log.debug("No update vertices to build vector index.")

The call to self.vid_index.to_index_file(self.index_dir) only happens if added_vids is non-empty.

So if you're only removing embeddings (i.e., no new vertices), the deletion is never persisted to disk.

---

Fix

removed_num = self.vid_index.remove(removed_vids)
self.vid_index.to_index_file(self.index_dir)  # <-- Add this line

---

Verification

• Remove only: works ✅
• Add only: works ✅
• No change: works ✅

Problem solved.

[Copilot]

Pull Request Overview

This PR ensures that vector removals in update_vid_embedding are persisted to disk by unconditionally writing the updated index after removals.

• Always call to_index_file after removing vids so deletions are saved.
• Retains existing behavior to write on additions.