What is an embedding in document search?

An embedding is a numerical vector that represents content so a system can compare meaning-related items. CNIL explains that vectorisation assigns numerical representations and that words close in meaning tend to have similar values in relevant components [1]. In a document-search project, the practical task is to decide what content to encode, how to count its inputs and how retrieved material will be checked.

Which OpenAI embedding model costs $0.02 per million tokens?

OpenAI lists text-embedding-3-small at $0.02 per million input tokens on the cited model page, as checked on 16 October 2026 [4]. That rate is a price for input tokens sent to that model, not a complete document-search bill. A project should count all submitted text, including repeated indexing when files change, before estimating this meter.

What is the text-embedding-3-small input limit?

The cited OpenAI embedding guide lists a maximum input of 8,192 tokens for text-embedding-3-small [5]. A long matter file may need to be divided into units that fit, but the limit alone does not prescribe the best division. Record token counts after the actual extraction and preparation step, then test whether the resulting search answers cite the right passages.

What does Gemini Embedding 2 text input cost?

Google's standard paid-tier price for Gemini Embedding 2 text input is $0.20 per million tokens on the cited price page, as checked on 16 October 2026 [6]. The number is a model-specific text-input rate. It should not be treated as a Vertex AI price, a PDF-processing quote or a total retrieval-system cost without separate evidence for those services.

How many PDF pages can Gemini Embedding 2 accept?

Google documents an 8,192-token overall input limit and one PDF per request with up to six pages for Gemini Embedding 2 [7]. Both constraints matter: a PDF under six pages could still exceed the token budget after extraction, while a longer PDF exceeds the stated page cap. Test representative files and record pages and tokens separately.

What Is an Embedding? OpenAI and Google Price Limits

Quick Answer: An embedding represents content as numerical values for search ¹. OpenAI text-embedding-3-small lists $0.02 per million tokens and an 8,192-token limit ⁴^,5. Gemini Embedding 2 lists $0.20 per million text tokens, an 8,192-token limit and one PDF of six pages per request ⁶^,7.

Summary in a mind map

What Is an Embedding? OpenAI and Google Price Limits
│
├─ Meaning · numerical similarity
│   ├─ Content becomes a vector of numerical values
│   ├─ Nearby vectors can identify candidate passages
│   └─ A candidate still needs source-document review
│
├─ OpenAI · one named model
│   ├─ text-embedding-3-small: $0.02 per million input tokens
│   └─ Maximum input: 8,192 tokens per request
│
├─ Google · price and PDF route
│   ├─ Gemini Embedding 2 text: $0.20 per million tokens
│   ├─ Overall input: 8,192 tokens
│   └─ One PDF per request, up to six pages
│
└─ Pilot · count actual submissions
    ├─ Log source format, pages and extracted tokens
    ├─ Count changed-file re-indexing separately
    └─ Check retrieved passages against file versions

What is an embedding for document search?

An embedding is a numerical representation of content. CNIL's explanation of vectorisation describes words associated with lists of values, or vectors, in a latent space ¹. Values are not hand-written labels. The model learns relationships from the context of words in its training material, and words that are semantically close can have similar values in some components ¹. For a search project, this means a query and candidate passages can be compared through their representations rather than through identical words alone.

Think of a law firm looking for a clause in many matter files. A person may ask for a concept using words unlike those in the source document. Numerical similarity can help identify a candidate passage. That candidate still needs review against the document itself: a nearby vector is a search signal, not a legal conclusion. The team must know which files were indexed and which passage was returned. The agent opportunity score is a separate question about whether such a task merits investment.

CNIL uses the familiar analogy of king - man + woman yielding a vector near queen, while cautioning that the resulting vectors are not exactly the same ¹. It also notes that vector length matters. The analogy illustrates how relationships can appear in a representation; it does not imply that a particular component is labelled “gender” or that arithmetic is a reliable search-quality test. The example should help the team understand the abstraction, not substitute for evaluating its own documents ¹.

OpenAI text-embedding-3-small price

OpenAI lists text-embedding-3-small at $0.02 per million input tokens on its cited model page, checked on 16 October 2026 ⁴. If a project submitted one million tokens once to that model, the stated input rate would imply $0.02 for that submission. This is multiplication of the published meter. It does not price document extraction, repeated indexing, a search database, user queries or any other system component.

The team should count the material it actually sends. A source document may be split into several passages for search, and a later revision may require those passages to be processed again. A budget that counts only the original file's length once will miss changed-file work. Keep the unit visible: the cited rate is per million input tokens, not per file, page or vector. AI agents for small businesses discusses the broader task design after the search substrate has been scoped.

OpenAI input tokens and file chunks

OpenAI's embedding guide lists an 8,192-token maximum input for text-embedding-3-small ⁵. Token length and page length are different measures. A short page full of unusual symbols can behave differently from a plain paragraph when tokenised. The practical check is to run the intended extraction and count the actual input tokens before submission. A file that exceeds the model limit must be handled through a deliberate preparation strategy, then tested for whether useful context survives.

An indexing worksheet can show file identifier, extracted-text token count, passage count and version. For a revised matter file, the version identifies which passages need another embedding call. These columns are a proposed accounting method, not a vendor promise. They answer a question the model limit creates: which inputs are within the documented ceiling and how much text will actually be charged at the model's token rate ⁴^,5.

The limit also constrains a single request, not the total number of documents a firm may hold ⁵. Do not describe an 8,192-token limit as an 8,192-token collection size. A collection may contain many separately prepared requests, but the number of requests, storage arrangement and retrieval quality depend on the project design. The distinction matters most when procurement staff hear a large file count and assume it directly determines model input capacity.

Gemini Embedding 2 price and limits

Google's cited pricing page lists $0.20 per million tokens for standard paid-tier Gemini Embedding 2 text input, checked on 16 October 2026 ⁶. The cited embedding guide lists an overall input limit of 8,192 tokens and one PDF per request with up to six pages ⁷. The price and constraints answer different questions. One is a text-token rate; the others control what a request can contain. None alone tells the firm the cost of preparing and indexing its entire document set.

Published item	OpenAI text-embedding-3-small	Google Gemini Embedding 2
Standard text-input rate	$0.02 per million tokens ⁴	$0.20 per million tokens ⁶
Input-token limit	8,192 ⁵	8,192 overall ⁷
PDF rule in these cited pages	No comparable PDF cap in the cited OpenAI limit ⁵	One PDF per request, up to six pages ⁷

The third row must be read carefully. An absence of a comparable PDF cap on one cited page is not evidence that any PDF can be sent directly through that model. It means the two pages document different input paths. A fair test prepares the same useful content for each route, records the processing step and checks the resulting answers. The model-rate comparison by itself is only one column in a system budget ⁴^,5^,6^,7.

Which counts belong in an indexing budget?

Start with the files a user actually needs to search. For each, record the source format, page count where applicable, extracted-text token count, model request count, and whether the file is a new version. For Google's PDF route, page count and the one-PDF rule are explicit admission checks ⁷. For both cited model limits, token count is an admission check ⁵^,7. For the text-input price comparison, tokens submitted are the direct chargeable unit ⁴^,6.

Record	Decision it supports
File and version	Which source text was indexed
Format and PDF pages	Whether the documented Google PDF route admits it ⁷
Extracted input tokens	Whether each model request fits its 8,192-token limit ⁵^,7; for a Google PDF, add 258 visual tokens per page to its text tokens ⁷
Re-indexed tokens	How much changed content is submitted again at the stated rate ⁴^,6

This is also a provenance record. A search result should be traceable to the underlying matter document rather than just to a vector identifier. The firm can compare a retrieved passage with the file version a lawyer expects. That check is especially important if the source files are updated or if an extraction step loses layout. The cited pricing pages do not guarantee that an embedding preserves the meaning of every clause ⁴^,6.

What an embedding does not settle

Numerical similarity is not proof that a generated answer is accurate, complete or safe to release. CNIL's vector explanation concerns representations, while the EDPB notes that information from training data, including personal data, can remain represented in model parameters and may be extractable under some circumstances ¹^,2. CNIL likewise explains that some models can contain extractable personal data from training, although an anonymous model may fall outside the GDPR ³. These are cautions about model status under EU data-protection analysis, not a claim that a US firm's document index is itself governed by that law.

Keep the distinction clear. A provider's training corpus, the firm's submitted matter documents, the vectors generated during indexing and the passages later shown to a user occupy different stages of the system, even though the same plain word “data” is often used for all four.

Those sources do not provide a vendor comparison of document-search security or a conclusion about a particular firm's data. They show why a numerical representation should not be called automatically anonymous. The law firm must assess its own files, contracts and deployment before moving confidential content into any service. It should also distinguish training-data questions from the separate operational question of what it sends for indexing ²^,3.

What to do next

Select a representative set of files and queries. Count the actual submitted tokens after extraction, note PDF pages, and map each request against the cited input limits. Run the same retrieval questions against the prepared content, then inspect whether the returned passages support a human answer. Keep model input rates in their own line and record unresolved preparation, storage and query charges separately. The comparison of agents and automation platforms and accounts payable example can help place retrieval within a larger workflow.