Lithi Documents & Search

Start with one cited outcome

For example, prepare cited chunks from POL-01, version A, for a returns question. Keep the 30-day rule and custom-item exclusion in one cited source region. If the page is unreadable or a qualification is missing, return that exception rather than filling the gap.

Your agent chooses the document evidence it needs to extract, retrieve or evaluate. It prepares permitted source versions and pages with clear citation and uncertainty rules. Lithi returns source-linked results and visible exceptions. Your agent checks the citations, qualifications and coverage against the source before using the result.

Your agent keeps the source question and acceptance decision. The Kit prepares the supported extraction or retrieval request; participating Macs process the bounded document evidence, and your agent checks citations and coverage.

PDFs, scans, forms, screenshots, and images belong to the Documents & Images information family. Choose the document evidence you need: extracted fields, cited chunks, embeddings, ranked results, or a retrieval evaluation. The operation determines the derivative and its evidence.

The operation is the exact supported capability. This guide explains preparation and review. Agent guidance does not answer from an unapproved source or grant access. The Documents & Search pack covers ocr-documents, extract-documents, review-layout, evidence-review, chunk-with-citations, embed-corpus, rerank-results, and evaluate-retrieval.

Conceptual task brief: Prepare cited chunks from POL-01, version A, for the returns question. The permitted evidence is the returns section and its page or heading locations. Keep the 30-day rule and custom-item exclusion together; retain unreadable or uncertain regions as exceptions. Return each chunk with its source identity and locator. Accept only when every cited claim can be checked against the source region and no qualification is lost.

The selected document operation and supported integration map that brief to the permitted request and returned evidence. A cited derivative still needs the customer’s review before a consequential decision.

Keep source evidence with every derivative

Freeze source files and versions, permitted pages, language or OCR profile, extraction schema, citation rules, chunk policy, embedding model and revision, vector dimensions and distance semantics, reranking contract, or evaluation corpus and relevance labels. Choose the operation that supplies the output you need.

1. Reconcile the page manifest before extraction. Preserve unreadable, protected, uncertain, and omitted pages as explicit outcomes.
2. Keep original text separate from display cleanup, summaries, and frontier corrections. Preserve page, region, table cell, row, column, heading, or section locators when the operation supplies them.
3. Keep chunks with document, version, section, page, and boundary context. A cited chunk shows where text came from; it does not show that a search index can find it.
4. For embeddings, keep stable chunk IDs, model revision, vector type and dimension, normalization, and distance semantics. Do not mix incompatible vectors.
5. For reranking or retrieval evaluation, preserve original candidate IDs and order, result ranks, relevance version, failures, and slice coverage.

Review citations and uncertainty

These are synthetic review targets, not universal filenames, output fields, or measured retrieval results. A schema-valid number can still be semantically wrong. Check the source region and its surrounding qualification.

An embedding artifact does not prove retrieval quality. OCR does not imply structured extraction, and extracted text does not imply layout reconstruction. A cited result still needs a reviewer to inspect the source before a consequential decision.

Recover source and vector failures

If OCR cannot read a page, retain that exception; do not report that the page contains no evidence. If a chunk boundary loses a negation or table header, revise the chunk policy and create a new derivative version. If vector dimension, model revision, or normalization differs, stop the import and resolve compatibility before joining results.

If delivery fails after an accepted artifact, retrieve it again. A corrected source, extraction schema, embedding model, or relevance set requires a new request and review. A cited chunk does not load an index, answer a question, or grant a database credential.

Worked example

Prepare cited chunks for retrieval

Freeze this small corpus and keep each qualification in its cited chunk:

Source POL-01, version A
Heading: Returns
Items may be returned within 30 days.
Custom items are excluded.

Source FAQ-01, version A
Heading: Packing
Keep the delivery label for reference.

POL-01/A/returns-1 must contain both the 30-day rule and the custom-item exclusion; FAQ-01/A/packing-1 must retain the delivery-label instruction. A chunk containing only the first returns sentence could answer “Can I return a custom item?” incorrectly. Preserve the document and heading rather than inventing page numbers for Markdown.

Compare published prices

You can see what it costs to read scanned pages with an AI model and with a document text service, using each provider's published list price. These tables compare published prices for the same work, not output quality.

The example reads 100,000 scanned pages:

• Each page is read as one 1,024-pixel image.
• About 500 tokens of text come back per page.
• Document text services bill per page, so each row uses its own unit.

Published list prices checked 5 October 2026.

Each provider counts tokens with its own tokenizer, so the same text can be a different number of tokens. Rows with a note count input with that provider's own published rule. Each service bills in its own unit. Lithi figures use the published example metering profile; a signed quote sets the real charge. These tables compare published prices for the same work, not output quality.

Use Image Analysis when image regions are the evidence and Translation when language handling changes the request. Continue with Result exceptions to keep conflicts, missing evidence, and delivery problems visible.