# Pinecone setup for search and RAG

> A vector database looks simple until the first re-index. We design the index, the metadata and the update path up front so you do not pay for the same mistake twice.

- Page: https://liftup.tech/platforms/pinecone-setup/
- Type: Platform
- Updated: 2026-11-20

## In short

Pinecone is a managed vector database used to store embeddings for semantic search and RAG chatbots. A sound setup fixes the embedding model and dimension first, uses namespaces to separate tenants, stores filterable metadata with every record, and plans how updates and deletions flow in. We also check whether a simpler store would do.

## Decisions we make before creating an index

Some Pinecone settings are painful to change later. The index dimension must match the embedding model, so changing the model means building a new index and re-embedding every record. We therefore settle these points first.

| Decision | Options | How we choose |
|---|---|---|
| Embedding model | Hosted general models, multilingual models | Test retrieval on your questions, not on a leaderboard |
| Dimension and metric | Set by the model, cosine is common | Follow the model's documentation |
| Deployment type | Serverless or pod-based | Serverless for most new projects, check current options |
| Tenant separation | Namespaces or separate indexes | Namespaces for many small tenants |
| Search type | Dense only, or dense plus sparse (hybrid) | Hybrid when exact terms such as part numbers matter |
| Metadata | Source, date, language, access group, version | Whatever you will need to filter on, decided now |

## Sizing math you can do yourself

Storage is easy to estimate. Each dimension of a standard vector is stored as a 32-bit number, which is 4 bytes. Take 50,000 chunks from your documents and an embedding model with 1,536 dimensions.

| Step | Calculation | Result |
|---|---|---|
| Bytes per vector | 1,536 x 4 | 6,144 bytes |
| Vectors only | 50,000 x 6,144 | about 307 MB |
| Metadata | 50,000 x about 1 KB | about 50 MB |
| Total, rough |  | about 360 MB |
| Same data at 3,072 dimensions | Vectors double | about 614 MB for vectors alone |

So a mid-sized document set is small in absolute terms, and the bill is usually driven by read and write volume and by how many separate indexes you run, not by raw storage. We list your expected reads and writes in the quote assumptions and tell you which plan terms to check. Our [AI chatbot cost calculator](https://liftup.tech/tools/ai-chatbot-cost-calculator/) helps you model the rest.

## What a setup usually contains

We have no published case study on Pinecone, so this is the typical scope and not a client story.

1. **Embedding test**: Compare two or three embedding models on 50 to 100 real questions and choose using retrieval hit rate.
2. **Index and namespace design**: Create the index, define namespaces per tenant or per document set, define the metadata fields.
3. **Ingestion job**: Batch upserts with stable record IDs, so re-running a file replaces its chunks rather than duplicating them.
4. **Query layer**: Filters for access group and date, optional hybrid search, optional reranking of the top results.
5. **Operations**: Deletion path for removed files, a re-index procedure, monitoring and a backup of the source text.

The application on top is usually a [RAG chatbot](https://liftup.tech/services/rag-chatbot-development/). If your documents are complex PDFs, see also our [LlamaIndex page](https://liftup.tech/platforms/llamaindex-development/).

## Gotchas

- Freshness. After an upsert, a record may take a short time to appear in queries. Do not build a flow that writes a record and instantly expects to read it back.
- Metadata limits. Each record can hold only a limited amount of metadata, so store the full text elsewhere and keep an ID in the record.
- Stable IDs. Random IDs make updates and deletes impossible to target. Derive IDs from file and chunk position.
- Access control. Filtering by user group must happen in the query itself. Hiding results after retrieval can leak content through the answer.
- Deleting data. Personal data removal requests need a way to find and delete every chunk of a person's file. Metadata makes that possible.
- Stale vectors after a model upgrade. Mixing old and new embeddings in one index produces poor results.

## Do you need Pinecone at all

Not always. If you already run PostgreSQL and have under a few hundred thousand chunks, the pgvector extension may be enough and removes a vendor. If you need very high query volume, multiple regions or low operations effort, a managed service earns its place. We recommend the smaller option when it fits, and the decision is part of the first week.

## Timeline

| Piece | Planning estimate |
|---|---|
| Embedding test and index design | 3 to 5 days |
| Ingestion and query layer for one document source | 1 to 2 weeks |
| Multi-tenant access rules, deletion and monitoring | 1 to 2 more weeks |

## Questions, answered

### Can we change the embedding model later?

Yes, but it means building a new index and re-embedding all content. Keep the cleaned source text so that step is a rerun and not a new project.

### Is Pinecone only for chatbots?

No. It also powers semantic site search, product recommendations, duplicate detection and similarity lookups. The index design differs for each.

### How do we keep tenants' data separate?

Use a namespace per tenant and require the namespace in every query. Test with a second tenant's account to prove nothing crosses over.

### What happens to our data if we leave?

Because we keep source text and the ingestion job, you can rebuild the index in another store. We document that path in the handover.

### Can you migrate us from another vector store?

Yes. We export the source text and metadata, re-embed if the model changes, and run both stores side by side for a short period while we compare results.

## Related

- [LlamaIndex development for document-heavy AI](https://liftup.tech/platforms/llamaindex-development/)
- [RAG chatbot development that cites your documents](https://liftup.tech/services/rag-chatbot-development/)
- [How much does a RAG chatbot cost in 2026?](https://liftup.tech/cost/rag-chatbot/)

---

Get a fixed quote: https://liftup.tech/contact/ · email rohit@liftup.tech · WhatsApp +91 77689 10234 (https://wa.me/917768910234) · reply within one working day.
Source page: https://liftup.tech/platforms/pinecone-setup/ (Liftup Technologies)
