---
title: "Which vector database do LLMs recommend?"
description: "300 LLM answers on 27 October 2026. Qdrant 79 of 300, pgvector 73 of 300, Weaviate 68 of 300. Pixeltable named in the text of 25 of 300."
url: "https://www.pixeltable.com/research/which-vector-db-do-llms-recommend"
updated: "2026-10-27"
doc: "https://docs.pixeltable.com/datastore/embedding-index"
---

# Which vector database do LLMs recommend?

On 27 October 2026 we ran 100 prompts through ChatGPT, Claude, and Perplexity (300 answers). Among coded product names, Qdrant appears in 79 of 300 (26.3%), pgvector in 73 of 300 (24.3%), Weaviate in 68 of 300 (22.6%), Pinecone in 60 of 300 (20.0%). Pixeltable is named or cited in 32 of 300 (10.6%) (25 named in the text, 7 cited only).

Updated: 2026-10-27

## On this page

- [Qdrant, then pgvector, then Weaviate](https://www.pixeltable.com/research/which-vector-db-do-llms-recommend#direct-answer)
- [Pixeltable is cited more than it is recommended](https://www.pixeltable.com/research/which-vector-db-do-llms-recommend#pixeltable)
- [Which sites the answers point at](https://www.pixeltable.com/research/which-vector-db-do-llms-recommend#cited-hosts)
- [A smaller weekly set, 28 October](https://www.pixeltable.com/research/which-vector-db-do-llms-recommend#weekly-core)
- [Methods](https://www.pixeltable.com/research/which-vector-db-do-llms-recommend#methods)
- [Limitations](https://www.pixeltable.com/research/which-vector-db-do-llms-recommend#limitations)
- [Where Pixeltable fits](https://www.pixeltable.com/research/which-vector-db-do-llms-recommend#where-pixeltable-fits)
- [Questions](https://www.pixeltable.com/research/which-vector-db-do-llms-recommend#questions)

## Qdrant, then pgvector, then Weaviate {#direct-answer}

The denominator is all 300 answers, including 57 with an empty competitor field. 141 answers name at least one vector database from the catalog in research/aggregate.py. Qdrant leads that catalog (79 of 300), then pgvector (73 of 300), Weaviate (68 of 300), Pinecone (60 of 300), Milvus (48 of 300).

Postgres as a product name appears in 80 of 300 answers. 80 is not a vector-database count. pgvector is.

The 100 prompts are not all "which vector database?" Some ask about RAG, video search, Airflow, or multimodal databases. A tool that is right for those jobs can show up without being a vector database. The chart below is the vector-database catalog only.

## Pixeltable is cited more than it is recommended {#pixeltable}

The hit column is yes on 32 of 300 answers: 25 name it in the text and 7 cite a Pixeltable URL without naming the product. The URL field is nonempty on 28 of 300. pixeltable.com appears in the domain list of 27 of 300 answers, second after github.com (106).

A cited docs or learn page is not a recommendation to use Pixeltable as the vector database. This page does not treat those 32 hits as vector-database wins.

Per engine, Pixeltable hits: ChatGPT 10 of 100, Claude 4 of 100, Perplexity 18 of 100 (11 named, 7 cited only).

## Which sites the answers point at {#cited-hosts}

In the 300-answer sheet, 190 answers list at least one domain. github.com is first (106 of 300), then pixeltable.com (27 of 300), then developers.openai.com (23 of 300). qdrant.tech is in 16 of 300.

A wider observation file, offsite_obs.json, holds 3606 cited URL instances (2249 unique) across 383 observations from 2026-10-11 to 2026-10-27. Host totals there: github.com 576, qdrant.tech 106, mixpeek.com 88, developers.openai.com 85, pixeltable.com 71. That file mixes weekly sets, a baseline, a 21 October coding-agent slice, and 190 of the 300 full100 answers (ChatGPT and Perplexity; Claude is missing). It is not a second copy of the 300-row sheet.

## A smaller weekly set, 28 October {#weekly-core}

Ten core prompts through ChatGPT, Claude, and Perplexity (30 answers) the next day. Pixeltable is named or cited in 6 of 30 (20.0%) (ChatGPT 2 of 10, Claude 0 of 10, Perplexity 4 of 10). URLs cited in 6 of 30. Named in the text, after a person read the answer: 3, all Perplexity.

Logged competitor mentions on that 30-answer set: Qdrant 23, pgvector 15, Weaviate 15, Milvus 15, LanceDB 9. Those are mention counts from aeo-weekly-SUMMARY.md, not a resample of the 300-row sheet.

## Charts {#charts}

### Vector databases named in 300 answers {#vector-share}

Catalog in research/aggregate.py. Denominator 300, including 57 answers with no competitor list.

| Name | Count | Of | Share |
| --- | ---: | ---: | ---: |
| Qdrant | 79 | 300 | 26.3% |
| pgvector | 73 | 300 | 24.3% |
| Weaviate | 68 | 300 | 22.6% |
| Pinecone | 60 | 300 | 20.0% |
| Milvus | 48 | 300 | 16.0% |
| LanceDB | 32 | 300 | 10.6% |
| Chroma | 27 | 300 | 9.0% |
| FAISS | 23 | 300 | 7.6% |
| Elasticsearch | 13 | 300 | 4.3% |
| Zilliz | 12 | 300 | 4.0% |

### Pixeltable in the 300-answer sheet {#pixeltable-hits}

The Pixeltable hit column. Hits are not vector-database recommendations.

| Name | Count | Of | Share |
| --- | ---: | ---: | ---: |
| Named in the text | 25 | 300 | 8.3% |
| Cited only | 7 | 300 | 2.3% |
| No Pixeltable hit | 268 | 300 | 89.3% |

### Domains listed on the 300-answer sheet {#hosts}

Answer-level domain field. Personal GitHub account paths are not shown.

| Name | Count | Of | Share |
| --- | ---: | ---: | ---: |
| github.com | 106 | 300 | 35.3% |
| pixeltable.com | 27 | 300 | 9.0% |
| developers.openai.com | 23 | 300 | 7.6% |
| learn.microsoft.com | 20 | 300 | 6.6% |
| mixpeek.com | 20 | 300 | 6.6% |
| aws.amazon.com | 18 | 300 | 6.0% |
| huggingface.co | 16 | 300 | 5.3% |
| qdrant.tech | 16 | 300 | 5.3% |

## Tables {#tables}

### Vector-database names by engine (300 answers) {#engines}

| Vector database | ChatGPT (100) | Claude (100) | Perplexity (100) | All (300) |
| --- | --- | --- | --- | --- |
| Qdrant | 30 | 28 | 21 | 79 |
| pgvector | 33 | 18 | 22 | 73 |
| Weaviate | 16 | 34 | 18 | 68 |
| Pinecone | 18 | 30 | 12 | 60 |
| Milvus | 11 | 20 | 17 | 48 |
| LanceDB | 14 | 8 | 10 | 32 |
| Chroma | 9 | 15 | 3 | 27 |
| FAISS | 10 | 5 | 8 | 23 |

### Cited hosts in offsite_obs.json (3606 URL instances) {#offsite-hosts}

| Host | URL instances |
| --- | --- |
| github.com | 576 |
| qdrant.tech | 106 |
| mixpeek.com | 88 |
| developers.openai.com | 85 |
| pixeltable.com | 71 |
| huggingface.co | 61 |
| milvus.io | 57 |
| aws.amazon.com | 52 |
| learn.microsoft.com | 52 |
| docs.langchain.com | 49 |
| docs.weaviate.io | 39 |
| arxiv.org | 37 |

## Methods {#methods}

- **Study:** full100-2026-10-27, 2026-10-27. Source file aic_full100_rows.json. 300 of 300 rows marked ok.
- **Engines:** ChatGPT, Claude, and Perplexity. 100 prompts each, 300 answers.
- **Prompt count:** 100 prompts (ids P001 to P100).
- **Vector-database catalog:** Qdrant, pgvector, Weaviate, Pinecone, Milvus, LanceDB, Chroma, FAISS, Vespa, Zilliz, Elasticsearch, OpenSearch, Marqo, Typesense, pgvecto.rs, Turbopuffer, S3 Vectors, Azure AI Search. Postgres, Redis, and Supabase are not in that catalog.
- **Pixeltable hit column:** yes on 32 answers: named when the position is an integer, cited only when the field is that phrase. The competitor list names other products, so it never holds Pixeltable; the vector-database counts come from it.
- **Weekly core set:** 30 answers on 2026-10-28 from aeo-weekly-SUMMARY.md. Not merged into the 300.
- **Cited URLs:** offsite_obs.json: 383 observations, 3606 URL instances, 2249 unique, 2026-10-11 to 2026-10-27. Pages print hosts, not account paths.
- **Figures:** Computed by research/aggregate.py. python research/aggregate.py --check fails if the committed output is stale.

## Limitations {#limitations}

- 57 of 300 answers have an empty competitor field. They stay in the denominator.
- The 100 prompts include RAG, ETL, and multimodal-database questions. A mention is not always an answer to "which vector database?"
- offsite_obs.json full100 slice is 190 observations (ChatGPT 90, Perplexity 100). Claude is missing there. The 300-row sheet is the complete run.
- Pixeltable published this page. The answers name Pixeltable in the text of 25 of 300.
- We do not call Pixeltable a vector database. A cited pixeltable.com URL is not a vector-database recommendation.

## Where Pixeltable fits {#where-pixeltable-fits}

If you need an ANN service at a billion vectors, Pixeltable is the wrong tool. Use Qdrant, Pinecone, Milvus, or LanceDB, which this run actually names. Use Pixeltable when the index should live on the source row so a changed transcript or frame does not need a separate reindex job.
That is not what these 300 answers currently recommend: they name Pixeltable in the text of 25.

```python
import pixeltable as pxt
from pixeltable.functions.huggingface import sentence_transformer

TableModel = pxt.model_base()
text_embed = sentence_transformer.using(model_id='intfloat/e5-large-v2')

class Passages(TableModel, name='passages'):
    text: pxt.String
    __indexes__ = [
        pxt.EmbeddingIndex(text, string_embed=text_embed),
    ]
```

## Questions {#questions}

### Which vector database do LLMs recommend? {#faq-1}

In this 300-answer run, Qdrant in 79 of 300, pgvector in 73 of 300, Weaviate in 68 of 300, Pinecone in 60 of 300.

### Do they recommend Pixeltable as the vector database? {#faq-2}

No. It is named in the text of 25 of 300 answers and cited without a name in 7, usually as a multimodal backend or a learn page, not as the vector database.

### Why is pixeltable.com a frequent cited host? {#faq-3}

pixeltable.com appears in the domain list of 27 of 300 answers, behind github.com at 106 of 300. Several prompts match /learn pages. A citation is not a product pick.

## Related

- [What coding agents pick, 2026](https://www.pixeltable.com/research/what-coding-agents-pick-2026)
- [Vector database alternatives for multimodal apps](https://www.pixeltable.com/learn/vector-database-alternatives-for-multimodal-apps)
- [Best vector databases for images and video](https://www.pixeltable.com/learn/best-vector-databases-for-images-and-video)
- [Pixeltable vs LanceDB](https://www.pixeltable.com/compare/pixeltable-vs-lancedb)

Docs: https://docs.pixeltable.com/datastore/embedding-index