Embeddings turn text into vectors, a foundational capability behind RAG, semantic search, deduplication, and clustering. Many people ask first: can I use it for free? The answer is yes, but the boundaries of "free" are broader than you might expect.

This article breaks down the common sources of free vectorization, explains the limits around quota, rate limits, privacy, dimensions, and quality, and then maps out scenarios that fit and do not fit, along with when to upgrade to a paid service. It does not quote specific quota figures, because free policies change quickly across platforms; always refer to the official pages.

Note: OpenAI and other names are trademarks of their respective owners; BGE, E5, GTE, Jina Embeddings, and similar names are models from their respective projects or communities. This article is a technical explanation only; free quotas, terms, and prices are subject to the official pages.

1. Where Free Embeddings Usually Come From

There are four common sources of free embeddings:

  • New-user trial credit on cloud platforms, usually with an expiration date and a call cap.
  • Free tiers of embedding APIs, rate-limited by request count or token count, suitable for small traffic.
  • Open-weight models that you download and deploy on your own servers or GPUs.
  • Small local models embedded in desktop applications or offline pipelines.

The cost structures of these four are completely different. The first two have someone else bearing the compute, at the price of quotas, rate limits, and contractual constraints; the latter two have you bearing the compute. The weights are free, but hardware, operations, and tuning are on you.

2. Where Free Tiers Usually Constrain You

A free tier is rarely as simple as a monthly token allowance. Common limits in practice include:

  1. Per-minute or per-day caps on requests and tokens, which batch jobs hit easily.
  2. A limit on batch size per request, requiring multiple calls when there are many documents.
  3. A maximum length per input, requiring long documents to be chunked first.
  4. Fixed or unavailable dimension choices, so changing models forces the index structure to change.
  5. No SLA on the free tier, with a higher chance of queueing or failures during peak hours.
  6. Data processing terms that may differ from the paid version, which is the most easily overlooked point.

So when evaluating a free option, do not look only at the allowance; consider rate limits, batching, dimensions, and terms together.

3. Open-Source Self-Hosting: The Weights Are Free, Not the Total Cost

Open-weight models (for example, public models from the BGE, E5, GTE, and Jina Embeddings families) can be downloaded and self-hosted, which suits teams whose data cannot leave the internal network. But "free" applies only to the weight files; actual costs include GPU or CPU inference resources, model selection and evaluation, service wrapping, monitoring, and upgrade maintenance.

Comparison Free API quota Open-source self-hosting
Upfront investment Low; sign up and use Requires hardware and deployment work
Marginal cost Constrained by rate limits and quota Mainly electricity and operations labor
Data location Data sent to the provider Data stays in your environment
Quality Depends on the provider's model Depends on model selection and tuning
Scalability Constrained by free quota Constrained by hardware capacity
Maintenance responsibility Provider Your own team

If the team has no GPU operations experience, the hidden costs of self-hosting often exceed expectations. Conversely, if data is sensitive and call volume is large and stable, self-hosting may be more controllable over the long term.

4. Run a Small Evaluation Before Deciding

The biggest risk of a free option is not spending money but choosing the wrong model and building an index with the wrong vectors. Follow these steps for a small evaluation that can be completed within two weeks:

  1. Select 50 to 100 representative texts from real business data, covering long and short lengths, languages, and business types.
  2. Prepare a set of queries with human-labeled relevant results as a retrieval baseline.
  3. Call the candidate free APIs and open-source models with the same data, recording vector dimensions and timing.
  4. Calculate hit rate, similarity distribution, and failure rate, focusing on Chinese and industry terminology.
  5. Estimate the total tokens for the full document set, the index rebuild time, and future query costs.

A short script can verify the interface and similarity calculations first:

from openai import OpenAI

client = OpenAI(
    api_key="sk-your-key",
    base_url="https://ai.zhongshangqihang.com/v1",
)

texts = ["What is the return policy?", "How do I request a refund?", "How do I get an invoice?"]
resp = client.embeddings.create(model="text-embedding-3-small", input=texts)
vectors = [item.embedding for item in resp.data]

def cosine(a, b):
    dot = sum(x * y for x, y in zip(a, b))
    norm = (sum(x * x for x in a) ** 0.5) * (sum(y * y for y in b) ** 0.5)
    return dot / norm

print(cosine(vectors[0], vectors[1]))

The model name in the example is for demonstration only; available models and dimensions are subject to model pricing and the in-site documentation.

5. Privacy and Compliance: The Most Easily Overlooked Cost of Free Tiers

Free services often trade looser data terms for usage volume. Before integrating, confirm at least three things: whether input data is retained, whether it may be used to improve or train models, and how deletion requests are handled. Terms may differ across platforms, regions, and free versus paid versions.

For scenarios involving customer contracts, personal information, financial data, and source code, do not treat a free public API as the default option. Use anonymized samples for functional validation first, and route production data to a paid service with a clear data processing agreement or to an internally deployed model.

6. Scenarios That Suit Free Vectorization

Suitable scenarios usually share these traits: non-sensitive data, small scale, tolerance for rate limits, and results used for validation rather than production decisions. Examples include personal learning projects, semantic search in a product prototype, feasibility validation for a small internal knowledge base, and offline batch clustering experiments.

Unsuitable scenarios are equally clear: external services that need a stable SLA, large daily document volumes, data subject to privacy or contractual obligations, and long-term fixed vector dimensions and model versions. In these cases, a free quota does not save money; it postpones risk until after launch.

7. The Upgrade Path from Free to Paid

A more reliable path is to validate quality with a free quota or open-source deployment first, then move production traffic to a paid API. Pay attention to three things during migration: whether the new model supports the same or mappable dimensions; whether old vectors can be retained, since changing models usually requires a full index rebuild; and whether the interface is compatible, since an OpenAI-compatible interface reduces code changes.

If you want to compare several models at once, you can test them through a unified gateway with a small volume first, then split keys and budgets by project instead of opening an account with each provider. Prices change, so refer to the model pricing page.

8. Frequently Asked Questions

How much free embedding quota is usually available?

It varies widely by platform. A common approach is a daily or monthly cap on requests or tokens, combined with per-minute rate limits. Specific numbers and expiration dates are subject to the official pages; do not plan capacity based on figures from hearsay.

Can free quota be used commercially?

It depends on the terms of service. Some free tiers allow commercial use but restrict data usage; others are limited to evaluation. Have legal counsel confirm before signing or launching; do not assume that usable means commercially usable.

Are open-source models always worse than commercial APIs?

Not necessarily. In Chinese retrieval scenarios, a suitable open-source model combined with good chunking and evaluation can meet many business needs. Gaps are more likely in long documents, multilingual content, and out-of-the-box stability, so test with your own data.

Will data on free tiers be used for training?

Policies differ across platforms and may differ between free and paid versions. If you cannot confirm, assume it may be used and do not submit sensitive data to public free interfaces.

Can old vectors be used after changing embedding models?

Usually they cannot be mixed directly. Vector spaces of different models are incompatible; after changing models you must re-embed all documents and update the model on the query side as well. This is why migration cost should be considered during model selection.

Next Steps

If you have already validated quality with a small sample, the next step is to route test traffic through a unified entry point: sign up for ZSQH AI Gateway, compare the dimensions, latency, and retrieval quality of different embedding models with a small allowance, and then decide on the production setup. The model list and prices are subject to model pricing.

Related reading: