Updating and Deleting Vectors

Data is never static. Your knowledge base articles get updated, users delete their profiles, and "Published" flags change to "Archived." In a traditional SQL database, UPDATE and DELETE are trivial operations. In a vector database, they are significantly more complex.

In this lesson, we will explore why updating a vector is different from updating metadata, how "Soft Deletes" work inside the HNSW graph, and how to perform Batch Deletions to keep your database clean and cost-effective.

1. Updating Vectors vs. Updating Metadata

In databases like Pinecone and Chroma, you can perform two types of updates:

Type A: Full Upsert (Changing the Vector)

If you rewrite an article, you must re-embed the text. To update this in the database, you perform a standard upsert using the same ID.

Impact: The database has to find the old vector, remove its links in the HNSW graph, insert the new vector, and build new links. This is a "Heavy" operation.

Type B: Metadata Update (Changing the Tags)

If you only want to change is_published: true to is_published: false, you don't need to touch the vector.

Impact: In modern vector databases, this is a "Lite" operation. It only updates the Metadata Store (Module 4) without touching the expensive Vector Index.

2. The Truth About Deletions (Tombstoning)

When you delete a vector from an HNSW graph, the database doesn't usually remove all the bits immediately. Doing so would "break" the graph structure for other users currently searching.

Instead, the database uses Tombstones:

The record is marked as "Deleted" in a bitset.
During a search, the query engine skips any ID marked as deleted.
Periodically, a Background Cleanup runner removes the tombstoned nodes and "re-links" the graph.

Performance Trap: If you delete 50% of your data but don't perform the cleanup, your searches will waste time "visiting" deleted nodes, leading to higher latency.

3. Strategies for Deletion

1. Delete by ID

The most common way. You know exactly what you want to remove. index.delete(ids=["doc_1", "doc_4"])

2. Delete by Filter (The Power Move)

You can delete everything that matches a metadata filter. index.delete(filter={"org_id": "closed_company_1"})

3. Clear the Collection

For testing environments. client.delete_collection("my_test_index")

graph TD
    A[Delete Command] --> B{Type?}
    B -- ID --> C[Direct ID Removal]
    B -- Filter --> D[Scan Metadata Index]
    D --> E[Collect IDs]
    E --> C
    C --> F[Tombstone Mark]
    F --> G[Background Cleanup]

4. Python Example: Updating and Deletion Patterns

Let's look at how to properly manage the state of an entry in Pinecone.

from pinecone import Pinecone

pc = Pinecone(api_key="your-api-key")
index = pc.Index("help-docs")

# 1. Update ONLY the Metadata
# Use update() instead of upsert() if you don't want to re-send the vector
index.update(
    id="doc_123",
    set_metadata={"status": "archived", "last_updated": 1704456000}
)

# 2. Delete with a Filter (Cleanup user data)
# Use this when a user cancels their account
index.delete(
    filter={"user_id": {"$eq": "user_abc"}}
)

# 3. Safe Deletion (Preventing Errors)
def safe_delete(index, doc_id):
    try:
        index.delete(ids=[doc_id])
    except Exception as e:
        print(f"Delete failed: {e}")

5. Metadata-Driven "Aging"

A common pattern in production is not to delete data, but to "Age" it using metadata.

Step 1: Add a created_at timestamp to all vectors.
Step 2: Every night, run a script that deletes all vectors where created_at < (today - 30 days).

This ensures your vector database doesn't grow infinitely and your search results stay "Fresh."

6. The Re-index Trigger

There comes a point where so many updates and deletions have happened that your ANN index (Recall) is suffering. This is when you trigger a Re-index.

Create a "New" collection.
Feed all fresh data into it.
Switch your app traffic to the new collection.
Delete the old, messy collection.

Summary and Key Takeaways

Maintenance is the key to a permanent AI system.

Prefer Metadata Updates: If the meaning (vector) hasn't changed, just update the tags.
Deletions are Async: Understand that "Deleted" data might still occupy space for a while (Tombstones).
Filter-based Deletes are the most powerful way to handle multi-tenancy.
Data Aging: Use timestamps to automatically prune old data and keep costs low.

In the next lesson, we move to Handling Versioned Data, exploring the complex problem of what to do when you have "Version 1" and "Version 2" of the same document in the same index.

Exercise: Cleanup Logic

You are building a "Chat History" feature.

Users can delete individual messages.
Users can delete their whole account.

Which deletion strategy would you use for "Delete individual message"?
Which strategy would you use for "Delete account"?
If you update the "Help Documentation" every Friday, should you:
- A) Upsert with the same IDs?
- B) Delete the old ones and Insert new IDs?
- C) Create a whole new collection?

Updating and Deleting Vectors: Maintaining Data Freshness