Gemini File Search is not entirely free. Within the Gemini API, Google states that file storage and generating embeddings at query time have no separate fee. Indexing documents can incur embedding charges, however, and questions that retrieve document passages consume the model's input and output tokens. The free tier has limits, so a step without a dedicated charge does not mean unlimited, cost-free RAG.
This explanation is based on Gemini Developer API documentation consulted on October 29, 2026. It is a documentation-based analysis: DTech did not upload PDFs, run API calls, or measure bills. Google's published prices may change.
What does File Search do with a PDF?
File Search is a document-retrieval tool connected to the Gemini API, not a subscription feature of the Gemini app. It uses RAG, short for retrieval-augmented generation: the system finds relevant passages in supplied files and passes them to a model as context for its answer. This is not the same as retraining Gemini on the PDF.
During preparation, the service imports the file, divides its contents into chunks, and builds a search index. An embedding is a numerical representation that helps find passages related to a question by meaning. At query time, the mechanism locates relevant parts of the index, and the model reads the retrieved context to formulate its answer. The documentation also supports citations and, for PDFs, page references when available, but these do not guarantee that every answer is accurate or that every page will be identified.
Which stages can cost money, and which have no separate fee?
| Stage | Documented rule | What to check |
|---|---|---|
| Upload, splitting, and preparation | The documentation does not list a separate PDF upload fee. | Processing produces data to index and counts toward service limits. |
| Indexing and embeddings | On paid tiers, indexing embeddings follow the price of the embedding model used. | Embedding model, tokens processed, and any reindexing. |
| Index storage | File Search has no separate storage charge. | Capacity limits still apply by tier; they are not token allowances. |
| Question embedding | Embeddings generated at query time have no separate fee. | This does not make the model's answer free. |
| Retrieved passages | Retrieved document tokens are billed as ordinary context on paid tiers. | The selected model's input rate and the size of passages used. |
| Generated answer | Input and output follow the compatible model's rates or free-tier allowances. | Model, modality, account tier, and tokens actually consumed. |
On the official pricing page consulted, File Search is listed as free of charge on the free tier, subject to applicable model conditions and quotas. On paid tiers, text embeddings with gemini-embedding-001 are listed at US$0.15 per 1 million input tokens. Google also documents gemini-embedding-2 for multimodal embeddings, with prices that vary by modality. The US$0.15 price should not automatically be applied to every kind of PDF or every model.
How much does it cost to index a PDF in practice?
There is no single price per file or megabyte. Embedding charges depend on the number of tokens actually processed and on the embedding model's modality. For example, if paid-tier text indexing consumes exactly 2 million billable tokens with gemini-embedding-001 at the rate above, the embedding component costs US$0.30. This is an illustrative calculation only, not a measured cost for an actual PDF or a bill forecast.
The full cost of a solution also depends on how often its document collection is indexed, the number of questions, the amount of material retrieved per answer, and the output tokens generated by the model. Reindexing changed material can add billable processing. Converting file megabytes directly into tokens without knowing how the file was interpreted produces an unreliable estimate.
What limits apply to PDFs and search stores?
File Search documentation sets a 100 MB limit per document. For the combined size of a project's search stores, it lists 1 GB on the free tier, 10 GB on Tier 1, 100 GB on Tier 2, and 1 TB on Tier 3. These are capacity limits, not storage prices.
Backend usage includes input data and stored embeddings. According to Google, the counted size is typically about three times the input file size, so a collection containing a nominal 1 GB of PDFs does not necessarily count as 1 GB of backend usage. The company recommends keeping each individual store under 20 GB to support retrieval latency.
There are also requests-per-minute, tokens-per-minute, and requests-per-day quotas, depending on the model and usage tier. These apply per project rather than per API key; Google AI Studio displays active restrictions. Storage capacity limits and request rate limits are different controls.
What changes for API users in Brazil?
Brazil is included in the list of regions supporting the Gemini API and Google AI Studio. That fact alone does not prove a particular model is enabled for a given account. Google ties paid tiers to billing configuration and account history. The public price list consulted states amounts in US dollars (USD); it would be misleading to convert them into a fixed Brazilian-real price or assume the final invoice currency and taxes without examining the relevant account and billing terms.
There is also a data-privacy distinction between tiers: Google's price table says that content submitted on the free tier may be used to improve its products, while describing a different condition for the paid tier. Before indexing contracts, personal data, or confidential documents, users should check current terms and the data handling appropriate to their use case.
How can you estimate spending without mixing up free features?
- Check the project's tier and billing arrangement in Google AI Studio.
- Verify that the generating model supports File Search, and identify the embedding model and the PDF's modality.
- Record the actual indexing tokens and apply the corresponding rate rather than using file size as a proxy.
- For questions, monitor input tokens, including retrieved passages, as well as model output tokens.
- Compare consumption, quotas, reindexing, and any taxes with the account's own billing dashboards.
In practical terms, storing PDFs and creating an embedding for each question add no File Search-specific fee, but indexing and generating answers can cost money. The final amount depends on the account, models, document modality, and actual consumption, rather than a supposed fixed monthly fee per PDF.
