embedding cost explosion error duplicate indexing fails

Not specified by the original author.

Vertex AI Vector Search: match distance measure and embedding dims or queries fail (indexconfigmismatch)

TL;DR

[Google Cloud networking reference (vertex-ai-index)]: your index config has to match your embeddings exactly. Distance measures are DOTPRODUCTDISTANCE, COSINEDISTANCE, or L2SQUARED_DISTANCE, and dims must line up (gemini-embedding-001 is 768). Pick shard size SMALL, MEDIUM, or LARGE.

Use this when you run into the situation in the title.

When not to use this skill: unrelated tasks. It covers only the procedure above.

Compatibility

The steps above apply to the commands named in them. This skill does not pin a version, so if a flag looks different on your machine, check your installed version's docs first.

Details

And if you want a public internet-accessible endpoint, you dont need VPC or Private Service Connect at all, it just hands you a domain name.

Context: Google Cloud networking reference (vertex-ai-index): index config must match your embeddings exactly. Distance measure types are DOTPRODUCTDISTANCE, COSINEDISTANCE, or L2SQUARED_DISTANCE, and the embedding dimensions must line up (gemini-embedding-001 and textembedding-gecko@003 are 768). Choose shard size SMALL, MEDIUM, or LARGE for the index. A public internet-accessible endpoint needs no VPC or Private Service Connect, it just generates a domain name for API calls.

Source: Google Cloud networking reference (vertex-ai-index): index config must match your embeddings exactly. Distance measure types are DOTPRODUCTDISTANCE, COSINEDISTANCE, or L2SQUARED_DISTANCE, and the embedding dimensions must line up (gemini-embedding-001 and textembedding-gecko@003 are 768). Choose shard size SMALL, MEDIUM, or LARGE for the index. A public internet-accessible endpoint needs no VPC or Private Service Connect, it just generates a domain name for API calls.

Verified resolution: [Google Cloud networking reference (vertex-ai-index)]: your index config has to match your embeddings exactly. Distance measures are DOTPRODUCTDISTANCE, COSINEDISTANCE, or L2SQUARED_DISTANCE, and dims must line up (gemini-embedding-001 is 768). Pick shard size SMALL, MEDIUM, or LARGE. And if you want a public internet-accessible endpoint, you dont need VPC or Private Service Connect at all, it just hands you a domain name.

Source URL: https://github.com/googlecloudplatform/cloud-networking-solutions/blob/HEAD/reference-architectures/networking-for-ai-inference/terraform/modules/vertex-ai-index/README.md

Not specified by the original author.

TL;DR: Dimension must equal the embedding model's output width and metric must suit the embedding type. Both are frozen at index creation; a mismatch fails every write, and the only fix is recreating the index. Create a new index with the right dimension, re-embed, re-upsert.

Dimension and metric: get them right at creation, they are frozen

The rule

dimension must equal the number of floats your embedding model outputs. metric must suit how you compare them:

• cosine: normalized embeddings (most text models, e.g. OpenAI text-embedding-3). The common default.
• dotproduct: unnormalized embeddings, or sparse-only indexes.
• euclidean: when you genuinely need L2 distance.

Why agents get this wrong

The failure mode is creating the index first and choosing the embedding model later, or swapping models mid-project. text-embedding-3-small outputs 1536 dims; text-embedding-3-large outputs 3072; many open models output 768 or 1024. Upserting 768-dim vectors into a 1536-dim index fails every record with a dimension-mismatch error. There is no alter-index; you delete and recreate, then re-upsert everything.

Checklist before create_index

1. Decide the embedding model first. Print len(embedding) once and use that number.
2. Pick the metric from the list above, not from habit. If your vectors are L2-normalized, cosine and dotproduct agree; if they are not, they do not.
3. If you need sparse vectors (keyword signals), set vector_type appropriately at creation too.
4. Name the index after the model and dimension (e.g. docs-ada-1536) so a future agent can see the pairing without reading code.

If you already messed it up

Do not try to "fix" the vectors to fit (truncating or padding silently destroys recall). Create a new index with the right dimension, re-embed, re-upsert. For large indexes use the bulk import path rather than re-upserting one by one.

When to use this

• This covers exactly what the title says: Pinecone.
• You are setting this up for the first time, or auditing an existing setup.
• You want the key gotchas in one place before you start.

When not to use this

• You are doing a different workflow with Pinecone; these steps are specific to the title above.
• You need the full reference docs; this is the short path, not the manual.

Compatibility

• Not pinned to a specific version; follows current Pinecone behavior.

Pinecone's create-an-index guide states dimension, metric, and vector type are set at creation and shows the supported metrics. The limits reference caps dense dimensionality at 20,000. The embedding-changed troubleshooting page covers float handling on upsert. Verified against live docs 2026-09-26.

Not specified by the original author.

TL;DR

You must also use the same embedding model that generated the indexed vectors; a query emb Pin one embedding model for a collection and store its name and dimension count alongside the index definition.

Steps

1. Pin one embedding model for a collection and store its name and dimension count alongside the index definition. Validate every query vector's length against the declared dimensions client-side before sending the aggregation; fail fast on mismatch. If you change embedding models, re-embed the collection and recreate the index; never mix models in one index.

When to use

You are seeing this: The queryVector you send at query time must be an array of exactly that many numbers. Use this skill when you run into "Match queryVector dimensions to the index and keep one embedding model per index".

When not to use

If your error message or symptom does not match what is described above, this is probably not your fix. Search for your exact error text instead of forcing this one to fit.

Versions

No specific versions are mentioned in the source material, so treat the fix as generally applicable and check the examples against whatever you have installed.

Why this happens

The original report does not dig into a root cause. It documents the symptom and the fix that resolved it.

MongoDB reference notes (grounded in the official Atlas Vector Search docs): the index definition declares the embedding field with its dimension count, for example { type: knnVector, dimensions: 1536, similarity: cosine }. The queryVector you send at query time must be an array of exactly that many numbers. You must also use the same embedding model that generated the indexed vectors; a query embedding from a different model has the right length but lives in a different vector space, so results are garbage with no error. Agents that swap embedding models mid-project or truncate vectors silently degrade recall.

Verified resolution: Pin one embedding model for a collection and store its name and dimension count alongside the index definition. Validate every query vector's length against the declared dimensions client-side before sending the aggregation; fail fast on mismatch. If you change embedding models, re-embed the collection and recreate the index; never mix models in one index.

Source: https://github.com/kaynetik/skills/blob/HEAD/ultimate-db/reference/mongodb.md