AI Context Splitter & Redactor

Paste the text you intend to use or load one supported UTF-8 text file.

Choose a maximum character count for each exported chunk.

Keep redaction enabled when detected patterns should be replaced before export, or deliberately disable it when the exact original text must be retained.

Analyze the context, review the finding categories and limitations, then download the numbered chunks and manifest.

The tool performs deterministic browser-side pattern scanning, optional replacement of detected spans, broad token estimation, and character-bounded text splitting.

It does not send the document to an AI model, summarize it, rewrite it, decide what is confidential, or approve it for disclosure.

The output is a preparation aid rather than a safety certification.

The scanner checks for a defined catalogue of recognizable text patterns. It does not use an AI model, provider API, breach database, or live credential-validation service.

High-confidence findings have a distinctive credential or private-key structure. Medium-confidence findings are more likely to overlap with ordinary text and require closer review.

Provider formats can change. A supported prefix does not prove that a token is active, and an unsupported or encoded secret can be missed.

A deterministic scanner is useful for recognizable formats such as configured credentials and personal-data patterns.

It cannot understand every secret format or the business meaning of ordinary sentences.

A customer name, unreleased price, internal incident, medical detail, or legal strategy can be sensitive without matching a distinctive technical pattern.

Manual review and organizational data-classification rules remain necessary.

When redaction is enabled, detected spans are replaced before the output is divided.

This prevents one detected value from being divided across two chunk files and ensures that the token estimate describes the exported text rather than the original source.

When redaction is disabled, every detected and undetected value remains in the chunks.

Replacing direct identifiers does not necessarily prevent a person, company, project, or event from being inferred from surrounding details.

Dates, locations, job titles, rare events, combinations of attributes, and linked external information can act as quasi-identifiers.

High-risk data sharing may require formal de-identification methods, risk measurement, governance, access controls, or a protected data environment rather than simple pattern replacement.

Documents, emails, web content, code comments, and retrieved records can contain instructions intended to manipulate an AI model or agent.

Sensitive-pattern redaction does not distinguish ordinary content from malicious instructions.

Treat every imported document as untrusted data, especially when the receiving AI can call tools, access accounts, write files, send messages, or perform external actions.

Use instruction and data separation, least-privilege tools, output validation, approval controls, and additional security testing in the receiving system.

Each chunk remains at or below the selected character ceiling.

The splitter first looks for a nearby paragraph break, then a line break, then other whitespace within the final portion of the available space.

When no suitable boundary exists, it uses the hard limit.

Windows line-ending pairs and supplementary Unicode code points are kept together.

Numbered filenames determine the intended order.

Joining the chunk contents in filename order reproduces the prepared output exactly.

The splitter does not add headings to the source text, repeat overlap between chunks, remove whitespace, or summarize content.

Users who need overlapping retrieval chunks, semantic embeddings, or recursive document parsing need a different workflow.

AI models process tokens rather than a universal character count.

Token boundaries vary with the model, tokenizer, language, code structure, punctuation, whitespace, and unusual identifiers.

The tool estimates a broad range of approximately one token per five characters through one token per two characters.

Use the target model's tokenizer when an exact count is required.

The correct size depends on the receiving model, existing conversation history, system instructions, attached tools, expected output length, and the task itself.

A model's context limit is not available exclusively to the document. Input instructions, previous messages, tool schemas, retrieved content, and generated output also consume tokens.

Smaller chunks can be easier to inspect and focus, while larger chunks preserve more surrounding context.

The file loader accepts TXT, Markdown, JSON, and CSV files up to 2 MiB.

Files are decoded with fatal UTF-8 validation. Invalid byte sequences stop loading rather than being silently replaced.

The tool does not guess Windows-1252, ISO-8859, UTF-16, or another legacy encoding.

Convert from the known source encoding with a controlled process and keep the original file.

A JSON or CSV file is loaded as plain text.

The tool does not validate JSON syntax, identify CSV delimiters, preserve spreadsheet column types, or split records semantically.

A hard boundary can occur inside a long JSON value, CSV record, code block, URL, or identifier when no suitable whitespace exists before the limit.

The ZIP contains numbered Markdown chunks, a JSON manifest, and a README.

The manifest records character counts, chunk filenames and sizes, the selected chunk ceiling, token range, redaction mode, and aggregated finding categories.

It intentionally omits the original filename and raw detected values to avoid adding unnecessary sensitive metadata to the pack.

Local preparation does not control what happens after a chunk is uploaded, pasted, synchronized, or sent through an API.

Review the receiving provider's account type, data retention, training controls, regional processing, administrator access, logging, connected tools, and deletion policies.

Use the minimum necessary data for the task and avoid sharing production secrets when a synthetic or reduced example would work.

A useful review separates technical pattern detection from authorization, data classification, prompt-injection risk, and provider controls.

The tool assists with only part of that process.

Text loading, pattern detection, redaction, splitting, token-range estimation, manifest creation, and ZIP compression run in the browser.

The implementation does not make an AI-model request or upload the source to an Olivez server.

Files are generated only after the download action is selected.