Summary: All three managed vector databases let you start a retrieval-augmented generation (RAG) prototype for $0 without a credit card, but their free tiers break in different places. Pinecone Starter caps monthly reads, writes and storage but has no idle timer in its limits docs. Qdrant's free cluster is a small single node that gets suspended after a week of no use. Weaviate's free cluster allows only 100,000 objects in one collection. Below I compare what each free tier actually allows, the first steps from the official quickstarts, and the errors you're most likely to hit.
What a RAG prototype needs from a vector database
A first RAG app usually does four things: split your documents into chunks, turn each chunk into an embedding, store the embeddings with the original text, and fetch the closest chunks when a user asks a question. The vector database handles steps three and four. Three practical questions decide which free tier fits:
- How many chunks will you store? Free tiers count this differently: Pinecone by storage size, Qdrant by RAM and disk, Weaviate by number of objects.
- Do you need a separate embedding API? All three now host embedding models, so you can skip OpenAI or Cohere keys while prototyping.
- Will the demo sit unused for weeks? This matters more than people expect, because two of the three free tiers suspend idle clusters.
Pinecone Starter: usage caps, built-in embeddings
Pinecone's Starter plan is free and covers its serverless database, Inference (hosted embedding and reranking models) and Assistant. Per Pinecone's pricing page and limits reference, Starter gives you:
- Up to 2 GB of serverless index storage per organization, 2M write units and 1M read units a month, and 1 GB of egress a month.
- Up to 5 serverless indexes in 1 project, 100 namespaces per index and 2 users per organization.
- 5M embedding tokens a month per model (for example
llama-text-embed-v2ormultilingual-e5-large) and 500 rerank requests a month onbge-reranker-v2-m3. - One location only: every Starter index must live in AWS
us-east-1.
Pinecone's pricing page includes its own illustrative estimate for a forum-answering RAG bot with 1024-dimension embeddings: roughly 130K category-scoped chats and 3K newly indexed messages a day on Starter. It notes that this counts database usage only, not embedding or reranking.
First steps (from the official quickstart)
- Create an account at app.pinecone.io and copy an API key.
- Install the SDK with
pip install --upgrade pineconeand setPINECONE_API_KEY. - Create an index with a dense vector field whose dimension matches your model.
llama-text-embed-v2outputs 1024 dimensions. - Embed your chunks with
pc.inference.embedusinginput_type="passage", in batches of no more than 96 inputs (that model's per-call limit), then upsert them along with the text. - Embed the user's question with the same model using
input_type="query"and search for the top matches.
If you use Claude Code, Pinecone also offers a plugin (claude plugin install pinecone, then /pinecone:quickstart) that does the schema, ingest and first query conversationally.
When you outgrow Starter, the next step is Builder at a flat $20 a month with no usage overages (10 GB storage, 5M write and 2M read units, any region), then Standard with a $50 monthly minimum and pay-as-you-go above it.
Qdrant Cloud Free: a real cluster, but it sleeps
Qdrant Cloud's Free tier is "free forever" and needs no credit card. Instead of usage units, you get a single-node cluster with 0.5 vCPU, 1 GB RAM and 4 GB disk. Qdrant's docs say this supports about 1 million vectors of 768 dimensions. Free clusters also include free Cloud Inference with selected models, so the cluster can create embeddings for you. The quickstart uses sentence-transformers/all-MiniLM-L6-v2, which produces 384-dimension vectors.
What you give up: there are no automatic backups (only manual snapshots through the API), version upgrades involve downtime, and the choice of cloud and region is limited. The big catch is idle handling. Unused free clusters are suspended after 1 week and deleted after 4 weeks if you don't reactivate them.
First steps (from the official Cloud Quickstart)
- Sign up at cloud.qdrant.io with email, Google or GitHub.
- Under Create a Free Cluster, name it, pick a provider and region, and click Create Free Cluster.
- Copy the API key when it's shown. It isn't displayed again.
pip install qdrant-client, then connect with your cluster URL, API key andcloud_inference=True.- Create a collection with a vector size that matches the model (384 for MiniLM) and cosine distance, upsert documents as text, and query with text. The embedding happens inside the request.
Paid Standard clusters are billed hourly for the vCPU, memory and storage you provision, plus backups and paid inference tokens. Qdrant points you to its pricing calculator rather than listing a flat entry price.
Weaviate Cloud Free: generous storage, one collection
Weaviate's Free plan costs $0 a month with no credit card. Each user gets one free cluster with 100,000 objects, 1 GB memory and 10 GB disk, 1 collection with up to 3 tenants, up to 2,000 embedding requests a day through Weaviate Embeddings, and 1,000 Query Agent requests a month. Hybrid search (vector plus BM25 keyword) is included, which helps when users search for product codes or exact phrases.
Free clusters run on AWS in a limited set of regions (the pricing table lists 2), and they're locked to the Cost Optimized (HFresh) index profile. To choose the Performance Optimized (HNSW) profile, you need a paid plan. Idle rules are similar to Qdrant's: after 7 days of inactivity the cluster is suspended with data preserved, and after 30 days of total inactivity it's permanently deleted. Weaviate sends a warning email 1 day before deletion.
First steps (from the official docs)
- In the Weaviate Cloud console, click + next to CLUSTER, choose Free, name the cluster and click Create cluster. Under Advanced configuration you can pick the AWS region and toggle auto-schema and CORS.
- Store the cluster URL and API key as
WEAVIATE_URLandWEAVIATE_API_KEY. - Connect with
weaviate.connect_to_weaviate_cloud(...)in the Python client. - Create a collection with the built-in
text2vec_weaviatevectorizer (no third-party key needed) and ingest your objects. - Query with
collection.query.hybrid(...).
The paid entry point is Flex, from $45 a month pay-as-you-go, with replication, backups (7-day retention) and up to 1,000 collections.
Side-by-side: free tiers and first paid step
| Pinecone Starter | Qdrant Cloud Free | Weaviate Cloud Free | |
|---|---|---|---|
| Price / card | Free | Free forever, no card | $0/mo, no card |
| Main free cap | 2 GB storage; 2M write units and 1M read units per month | 1 node: 0.5 vCPU, 1 GB RAM, 4 GB disk (about 1M 768-dim vectors) | 100,000 objects, 1 GB memory, 10 GB disk |
| Structure limits | 5 indexes, 100 namespaces per index, 1 project | Single node, no high availability | 1 collection, up to 3 tenants |
| Hosted embeddings | 5M tokens/month per model | Free models included (for example MiniLM, 384-dim) | 2,000 requests/day |
| Idle rule | No idle suspension found in limits docs | Suspended after 1 week unused; deleted after 4 weeks | Suspended after 7 days; deleted after 30 days (email 1 day before) |
| Region | AWS us-east-1 only | Limited choice | AWS, 2 regions |
| First paid step | Builder, $20/month flat | Standard, usage-based (hourly) | Flex, from $45/month |
Which one should you start with?
Some back-of-envelope math helps. A 1024-dimension float vector is about 4 KB before metadata, so Pinecone's 2 GB holds a few hundred thousand chunks at most, fewer once you store the chunk text. Weaviate's 100,000-object cap is easier to reason about: if each document splits into about 20 chunks, you can index around 5,000 documents. Qdrant's 1 GB of RAM is the tightest resource, but its own estimate of about 1M 768-dimension vectors means a typical help-center or course-notes corpus fits easily.
- Pick Pinecone Starter if your demo needs to stay reachable for weeks without traffic (a portfolio project or a stakeholder demo) and you want hosted embeddings and reranking from one API key.
- Pick Qdrant Free if you want a real Qdrant instance whose code carries over directly to self-hosted or paid clusters, with heavy metadata filtering, and you'll keep using it at least weekly.
- Pick Weaviate Free if you want hybrid keyword-plus-vector search out of the box and your corpus fits in one collection of 100,000 objects.
Editor's note
Good fit: developers building a first RAG chatbot, semantic search or internal knowledge tool who want managed hosting without a card on file.
Not a good fit: anything production-facing on a free tier. None of the three free tiers includes an uptime SLA, Qdrant and Weaviate free clusters have no backups, and both suspend idle clusters.
Common errors and fixes:
- Pinecone
403 QUOTA_EXCEEDED: you've hit an object limit, such as a 6th index or a 2nd project. Use namespaces inside one index instead of extra indexes. - Pinecone
429 TOO_MANY_REQUESTS: either a per-second limit (back off and retry) or the monthly read or write unit cap. The monthly one won't clear until next month unless you upgrade. - Pinecone index in the wrong region: Starter only allows AWS
us-east-1, and downgrading from a paid plan deletes indexes you can't keep there. - Qdrant 403 right after creating a key: Qdrant says a new key can return 403 for a few seconds. Retry a handful of times before assuming the key is wrong. Also don't mix up a Database API key (for one cluster) with a Cloud Management Key (for
qcloud). - Qdrant filter rejected: new Cloud collections run in strict mode, which rejects filters on fields without a payload index. Create the payload index before loading data.
- Dimension mismatch on any of them: vector size is fixed when the collection or index is created. Changing embedding models means creating a new one and re-ingesting.
- Demo is down after a holiday: on Qdrant or Weaviate it was probably suspended for inactivity. Reactivate it from the console before the deletion deadline.
See the official screenshots at the end of this post for the plan cards and lifecycle notes I checked.
FAQ
Do I need an OpenAI key to try these?
No. Pinecone Inference, Qdrant Cloud Inference and Weaviate Embeddings all host embedding models within their free allowances. You'll still need an LLM to generate the final answer in a full RAG app, but retrieval works without one.
Can I move my data when I upgrade?
Weaviate says you can upgrade a free cluster to a paid plan without losing data, and Qdrant says you can upgrade to a Standard cluster. On Pinecone, upgrading is a plan change for your whole organization under Settings > Billing > Plans, and the higher quotas apply immediately. Builder is card-only and has no usage overages.
Which free tier lasts longest without me touching it?
Pinecone Starter. Its limits are monthly usage caps rather than an idle timer. Qdrant suspends unused free clusters after 1 week and Weaviate after 7 days, and both delete them after about a month.
Related on lightmind24: Dify Playbook: Turn Your Help Docs Into a Support Chatbot, LM Studio Tutorial: Run a Local LLM and Chat With Your Files, and OpenRouter Free Models: Limits, Fees and When to Pay.
Start with the free tier whose limit you're least likely to hit, and set a weekly reminder if that tier is one that falls asleep.
Images: screenshots of the public pricing pages and documentation of Pinecone (pinecone.io), Qdrant (qdrant.tech) and Weaviate (weaviate.io, docs.weaviate.io), captured by lightmind24.
Last verified: 2026-11-01
Sources: Pinecone Pricing, Pinecone Object Limits, Pinecone Rate Limits, Pinecone Quickstart, Pinecone: Ingest Your Own Files, Pinecone: Downgrade Your Plan, Qdrant Pricing, Qdrant: Creating a Cloud Cluster, Qdrant Cloud Quickstart, Qdrant llms.txt, Weaviate Pricing, Weaviate Cloud: Create a Cluster, Weaviate llms.txt.
댓글
댓글 쓰기