Pinecone−63 Strong negative

Operator Insights

What are teams learning from building on Pinecone and running it in production?

4 accounts: 1 failure, incident or rollback, 1 left it or turned it down, 1 switched and 1 adopted it

Failures, incidents and rollbacks1

Mohammed MohsinFounding EngineerOmi
Omi's account deletion flow, which removes each user's screen-activity vectors keyed by user id from Pinecone.
GitHub, BasedHardware/omi, published Aug 31, 2026Read the pull request (opens GitHub, BasedHardware/omi in a new tab)
Observation
In a production incident at Omi, account deletions failed on one step because one call site sent Pinecone deletes of more than 1,000 ids, which Pinecone rejects with a 400.
Takeaway
Chunk every bulk delete and upsert to Pinecone's per-request limits, and test deletion on accounts with more than 1,000 records.
What happened
Omi's pull request, opened Aug 31, 2026, says account deletions had been failing in production on one step: Pinecone rejects a delete carrying more than 1,000 ids with a 400, so an account with more than 1,000 screenshots could not be deleted at all. Omi's own diagnosis is that this one call site was missed when deletes were chunked elsewhere. The fix chunks the call and was merged Sep 1, 2026.
- Public engineering write-up (Omi's own pull request)
- Cause stated by the team
- Opened Aug 31, 2026
Deployment details
Omi's account deletion flow, which removes each user's screen-activity vectors keyed by user id from Pinecone.
- Public engineering write-up (Omi's own pull request)
- Cause stated by the team
- Opened Aug 31, 2026
Team
Omi
Duration
Fixed Sep 1, 2026
Workload
Deleting a user's screen-activity vectors keyed by user id
Measurement
Deletes above 1,000 ids rejected with a 400

Left it or turned it down1

Jeffrey IpCo-founderConfident AI
Confident AI, which builds open-source LLM evaluation infrastructure, clustering LLM observability log data in production.
Confident AI blog, published Oct 29, 2023Read the post (opens Confident AI blog in a new tab)
Observation
Confident AI dropped Pinecone for pgvector after weeks of experimentation, finding that Pinecone has excellent features for POC projects but needs substantial effort to keep scalable.
Takeaway
If your searchable data already lives in one database, check how much metadata each record needs before choosing a separate vector store.
What happened
Confident AI used Pinecone to cluster LLM observability log data in production, then after weeks of experimentation replaced it entirely with pgvector. The post says Pinecone has excellent features for POC projects but requires substantial effort to maintain a scalable and performant search infrastructure, and that Pinecone limited metadata to 40 KB per vector, which meant extra queries to the main data store.
- Public organization write-up (co-founder's post on the company blog)
- Outcome stated
- Published Oct 29, 2023
Deployment details
Confident AI, which builds open-source LLM evaluation infrastructure, clustering LLM observability log data in production.
- Public organization write-up (co-founder's post on the company blog)
- Outcome stated
- Published Oct 29, 2023
Team
Confident AI
Duration
Weeks of experimentation
Workload
Clustering LLM observability log data in production

Switched to or from it1

Hanna PaasivirtaAI EngineerOpenFn
Apollo, OpenFn's AI services backend for its workflow automation platform.
GitHub, OpenFn/apollo, published Nov 10, 2025Read the issue (opens GitHub, OpenFn/apollo in a new tab)
Observation
OpenFn is moving its embeddings from Pinecone to Postgres so that it manages one database and gains keyword search. Postgres support merged on Sep 18, 2026 while production still ran on Pinecone.
Takeaway
Teams that already run Postgres weigh one less database against Pinecone's managed search. At OpenFn the move had not reached production more than ten months after the decision.
What happened
On Nov 10, 2025 OpenFn opened an issue saying that ideally it would manage only one database, so it would replace Pinecone with Postgres for its embeddings and add keyword search. A pull request adding Postgres support merged on Sep 18, 2026. Issues dated Sep 24 and Sep 29, 2026 say production was still on the release using the Pinecone embeddings, so the switch was not complete on Sep 29, 2026.
- Public engineering write-up (issues and pull request in OpenFn's own repository)
- Switch in progress
- Opened Nov 10, 2025
Deployment details
Apollo, OpenFn's AI services backend for its workflow automation platform.
- Public engineering write-up (issues and pull request in OpenFn's own repository)
- Switch in progress
- Opened Nov 10, 2025
Team
OpenFn
Duration
Nov 2025 to Sep 2026, ongoing
Workload
Embeddings for Apollo, OpenFn's AI services

Adopted it1

ShortwaveOct 17, 2023Shortwave blogShortwave stored its email embeddings in Pinecone, namespaced per user, and named that namespacing as one of the main reasons it chose Pinecone. Its AI answered almost all questions within 3 to 5 seconds end to end.
Hiranya JayathilakaEngineerShortwave
Shortwave's email AI assistant, which answers questions over each user's own mailbox.
Shortwave blog, published Oct 17, 2023Read the write-up (opens Shortwave blog in a new tab)
Observation
Shortwave stored its email embeddings in Pinecone, namespaced per user, and named that namespacing as one of the main reasons it chose Pinecone. Its AI answered almost all questions within 3 to 5 seconds end to end.
Takeaway
Per-user namespaces fit an assistant over each user's own data, and the end-to-end time Shortwave reports covers the whole pipeline, of which Pinecone is one part.
What happened
Shortwave embeds emails and stores the embeddings in Pinecone, namespaced per user, and says namespacing is one of the main reasons it chose Pinecone. For other emails it fetches the stored vectors from Pinecone and computes cosine similarity locally. Shortwave reports that despite the system's complexity it answers almost all questions within 3 to 5 seconds, end to end.
Public engineering write-up
Measured results reported
Published Oct 17, 2023
Deployment details
Shortwave's email AI assistant, which answers questions over each user's own mailbox.
Public engineering write-up
Measured results reported
Published Oct 17, 2023
Team
Shortwave
Workload
Email embeddings for an AI assistant, one namespace per user
Measurement
Almost all questions answered within 3 to 5 seconds end to end (Oct 17, 2023)

Failures, incidents and rollbacks1

In a production incident at Omi, account deletions failed on one step because one call site sent Pinecone deletes of more than 1,000 ids, which Pinecone rejects with a 400.

Left it or turned it down1

Confident AI dropped Pinecone for pgvector after weeks of experimentation, finding that Pinecone has excellent features for POC projects but needs substantial effort to keep scalable.

Switched to or from it1

OpenFn is moving its embeddings from Pinecone to Postgres so that it manages one database and gains keyword search. Postgres support merged on Sep 18, 2026 while production still ran on Pinecone.

Adopted it1

Shortwave stored its email embeddings in Pinecone, namespaced per user, and named that namespacing as one of the main reasons it chose Pinecone. Its AI answered almost all questions within 3 to 5 seconds end to end.