Skip to main content

Search and AI

OpenKeep includes two related capabilities:

  • archive search
  • AI-assisted summaries and answers

They work together, but they are not the same thing.

For the broader archive workflow, see Core Workflows.

Search Basics​

Open Chat and ask a question.

OpenKeep runs hybrid search, which combines:

  • keyword matching
  • semantic similarity

Search is answer-first. Instead of only showing a traditional result list, OpenKeep tries to return the most useful archive answer for the question you asked.

This helps with both exact terms and more natural-language questions.

For some questions, OpenKeep also uses structured archive state instead of relying only on free-text retrieval. This is especially important for operational questions about open work, review queues, and deadlines.

How Search Results Appear​

Depending on the question, search can return one of two main answer styles.

Semantic Answers​

For exploratory or knowledge-style questions, OpenKeep returns:

  • a generated answer
  • linked citations back to source documents
  • supporting source cards

This path is best for questions like:

  • What does the archive say about a particular supplier?
  • Which document mentions a contract end date?
  • What amount is mentioned in this letter?

Structured Answers​

For operational questions, OpenKeep may answer directly from structured archive fields such as due date, expiry date, review status, and task completion state.

These answers can show:

  • a short summary
  • linked item cards
  • counts
  • totals when applicable
  • due dates, expiry dates, review reasons, or action labels

This path is best for questions like:

  • Which invoices are still open this month?
  • Which documents still need review?
  • Which contracts expire soon?

Ask Your Archive​

Chat is a conversation, not a one-shot search box. Type a question in the composer at the bottom and press Enter; Shift+Enter inserts a newline.

Examples:

  • When is the invoice due?
  • Which document mentions a contract end date?
  • What does the archive say about a particular supplier?
  • Which invoices are still open this month?
  • Which documents still need review?
  • Which contracts expire soon?

When successful, OpenKeep shows:

  • a generated answer
  • linked citations back to source documents for semantic answers
  • structured item cards for operational answers

If the evidence is too weak, OpenKeep may refuse to answer confidently.

Conversations​

Every question you ask joins the conversation open in the left rail. Use New conversation to start a fresh thread, and pick an earlier one from Recent to reread it.

Conversations are stored in your browser, not on the server, so they stay on the machine you asked them from and are not part of an archive export.

In the desktop app, every archive profile has a separate browser partition. Its conversations and recent searches persist when you return to that profile, but they never appear in a different profile and are removed with that profile's local state.

Citations​

Click a citation marker in an answer to open the quoted passage beside it. From there you can copy the quote or jump straight to that place in the source document.

Only extracted text is ever sent to the chat provider — never the original file.

On mobile the same markers open the cited document at the cited page with the quoted passage highlighted above it. Sources that are not in the device's offline copy are marked as such while the archive is unreachable. See Mobile App.

How to Judge an AI Answer​

Treat AI answers as a convenience layer over your documents, not as the final authority.

Always verify:

  • the cited document
  • the quoted excerpt
  • important dates, amounts, and obligations

For structured answers, verification may mean opening the linked document card and checking the extracted fields that were used for the answer.

This is especially important for invoices, contracts, tax documents, and legal records.

Document Summary​

Each document detail page includes an AI summary section.

OpenKeep tries to generate a concise explanation of what the document is about.

You can:

  • let it generate automatically
  • regenerate it manually
  • see provider and model information when available

The web and desktop clients show summary text as it streams. Leaving the document, closing the desktop window, or switching desktop profiles cancels the prior stream so an incomplete result is not presented as finished.

Ask Questions About a Single Document​

Each document detail page also includes a Q&A section.

Use it to ask focused questions about one document, for example:

  • What is the due date?
  • What amount is listed?
  • Who is the issuer?
  • What obligations are described in this contract?

OpenKeep streams the answer and stores a persisted Q&A history for that document. Each question is answered on its own — ask complete questions rather than referring back to a previous answer.

When source excerpts are shown​

For long documents OpenKeep selects the passages most relevant to your question and lists them as source cards you can open.

No source cards appear in two cases, which the answer text itself distinguishes:

  • Short documents (most letters and invoices): the complete document text is sent to the model, so separate excerpts would just repeat it. The answer references pages inline, for example "on page 2".
  • Long documents without a usable search index: only the beginning of the document is used. The answer says so explicitly. Treat such answers with care — a fact further back in the document may have been missed. Configuring an embedding provider and letting the document finish indexing restores full retrieval.

Clearing Q&A History​

The Q&A section includes Clear history.

Use this when you want to remove earlier question-and-answer entries for that document.

Intelligence Tab on a Document​

The Intelligence tab shows how OpenKeep interpreted the document.

Depending on the document, it may include:

  • routing result and confidence
  • generated summary
  • extracted type-specific fields
  • field confidence
  • provenance snippets
  • tagging and correspondent resolution
  • validation warnings or errors
  • pipeline metadata

This is most useful when you want to understand why a document was classified or flagged a certain way.

When AI Features May Be Limited​

AI-generated summaries and answers depend on configured providers and on document quality.

You may see limited or missing AI output when:

  • no chat provider is configured
  • OCR quality is poor
  • the document has very little usable text
  • the archive does not contain enough evidence for the question

Good Search Habits​

  • start with plain-language questions
  • ask operational questions explicitly when you care about current state, for example still open, due this month, pending review, or expires soon
  • try a narrower query if you get too many results
  • use document search and document-level Q&A together
  • check citations before acting on an answer

Next Step​

Continue with Review and Corrections to handle uncertain extraction results and fix metadata safely.

For admin and provider setup topics, continue with Settings and Admin after that.

Ask by Life Domain​

The chat understands the archive's categories. Questions like "Was gebe ich für Versicherungen aus?" or "show everything around housing from 2025" filter and aggregate through the category assignment instead of guessing from document text — sums come back grounded, per currency, and a category the archive does not have is reported honestly rather than invented. Categories compose with every other filter (year, correspondent, type) in the same question.