Free RAG API: what AIVAX Free includes for embeddings, reranking and OCR

Lists of free embedding APIs, like the one on Free LLM API Hub, compare embedding models. A RAG prototype also has to turn PDFs and scanned pages into text, store and search the chunks, and often reorder the candidates before a model writes the answer, and assembling those from separate free tiers means separate keys and quotas.

AIVAX Free covers four of those steps under one account, each with its own daily allowance: embeddings for inserting into and searching AIVAX collections, Reflex reranking, Julia-1 semantic decisions, and Fetch/OCR extraction. There is no standalone embeddings endpoint; the embedding allowance applies to collection insertion and search. The model that writes the final answer is always billed, and the account needs a positive balance even when everything it does falls inside the allowances. Per day, they are 10,000 embedding tokens, 2,000 Reflex tokens, 20,000 Julia-1 input tokens and 1,000 Fetch/OCR processing units.

Four allowances, one per service

The allowances are separate and do not transfer: unused OCR capacity cannot pay for embeddings. Plans and limits defines what each covers.

The RAG embedding allowance pays for document insertion and for search queries from one shared pool. A query embedding served from cache does not consume it. Media processing, document extraction, summarization, classification and segmentation are outside it. The Reflex allowance pays for Reflex reranking after a collection search or through the standalone rerank endpoint, and it counts cached and uncached input alike; the third-party rerankers compared in best reranker for RAG are never covered. The decision allowance covers semantic decisions served by @supersonic-labs/julia-1 only; other decision models are always billed. The Fetch/OCR allowance pays for extraction from web pages, documents and images, measured in processing units, but not for the optional schema-guided JSON conversion.

Paid plans multiply each allowance by a different factor: Pro gives 15× the embeddings, 4× Reflex, 2× Julia-1 and 8× Fetch/OCR, and Max and Ultimate go higher. The pricing page has the current table.

The daily amounts are measured in tokens and PUs

Each allowance counts what its service already meters. The 10,000 embedding tokens are tokens sent to the embedder, insertion and search combined. The 2,000 Reflex tokens are input tokens, cached input included. The 20,000 Julia-1 tokens cover input only, so a decision's output tokens are always billed. The 1,000 Fetch/OCR units are the same processing units each extraction reports.

The pricing page turns the embedding allowance into about 400 short queries a day, assuming 20 tokens per query, no insertions that day and some headroom. It gives no request estimate for Reflex or Julia-1, and for planning I would not trust one either, because the size of each item decides more than the number of calls.

Coverage is decided per metered item, and each item is either fully included or billed in full. An item is a single document's embedding, one query-term embedding, one reranking call, one decision call's input, or one extraction. The allowances tolerate a 10% margin above their base; an item that would cross that margin is billed normally and leaves the allowance untouched. A large item can therefore be billed while smaller ones later in the day are still covered.

The token and PU estimates published with the limits show which items are large. A 1,500-character chunk is about 350 to 400 embedding tokens. Reranking 50 candidates of about 300 tokens each is roughly 15,000 tokens in one call. A native-text PDF costs about 1 PU, while a ten-page scanned PDF costs 400 to 500, because OCR drives the total.

Against the Free amounts, the embedding allowance holds about 25 to 30 chunks of that size a day, before any query. A 50-candidate rerank is more than seven times the Reflex allowance, so it is always billed on Free; a rerank stays covered only while the candidates fit in about 2,000 tokens, roughly six passages of 300 tokens. Two scanned ten-page PDFs use up the OCR allowance, while native-text PDFs barely touch it. A prototype that indexes clean text in small batches and reranks a short list can stay inside the allowances; one that indexes scanned archives starts paying for OCR after the second file.

What exceeds an allowance is billed to your balance at product rates. Uncovered OCR costs $0.15 per thousand processing units on Free, so a scanned ten-page PDF of 400 to 500 units is about $0.06 to $0.08. Reflex input costs $0.015 per million tokens, and Free usage carries a 25% markup, so reranking 50 candidates of about 300 tokens each, roughly 15,000 tokens, is about $0.0003. The pricing documentation links the current rate of each product, embeddings included.

Allowances reset daily at midnight in the server's local time; the docs do not name the time zone, so the usage indicators are the reliable way to see when yours resets. They never bypass rate limits, and the decisions API requires a positive balance even for a zero-priced model.

Inference on Free runs on low-price models

LLM subscription coverage is currently disabled on every plan, so the answer call is metered like any other inference. On Free, inference is limited to low-price models and billed at 1.25 times the model's base price, and it is capped at 20 requests per minute, 500 per day and 64,000 input tokens of context. The pricing documentation explains how anything outside the allowances is charged, and how to get the cost of each LLM API call shows how to read what a request actually charged.

Free plan request limits for a RAG prototype

An allowance that runs out turns into billing. A request limit that runs out turns into an error until the window resets. These are the Free values from Plans and limits:

Fetch requests share the 30 general service actions per day with other services, so send several URLs or files in one contents array instead of one request each. The 500 insertions a day is a request ceiling, not what the allowance pays for: at 350 to 400 tokens per chunk, only the first couple of dozen insertions fit in the 10,000 embedding tokens, and they draw from the same pool your search queries need.

Testing a RAG pipeline inside the Free plan

We test one stage at a time, since each stage draws on a different allowance and a problem in one is easy to blame on another:

1. Extract a handful of representative sources with Fetch and check the extraction quality before spending embeddings on bad text.
2. Insert a small collection, run real questions and read the returned passages before adding generation.
3. Rerank only if the right passages are retrieved but ranked too low; Do I need a reranker for RAG? explains how to tell.
4. Use a Julia-1 decision where a step needs a category or a score instead of prose.
5. Add the answer model last and measure its cost on its own.

When the prototype outgrows an allowance, the per-item billing means individual items start to be charged while the rest of the day's usage stays covered; the request limits above do not soften the same way. Create an AIVAX account to try it with your own documents.

Last verified on 9 November 2026 against the AIVAX source, the Plans and limits documentation and the pricing page.