Search by keyword and by meaning
Retiform searches everything your team has uploaded in one query. You don't choose between a keyword search and a search by meaning, and neither does your agent. By default, a search runs both and ranks the results together.
Why one search, not two
People who work through a large set of documents ask two kinds of question without thinking about which is which.
- Exact things. A reference number, a name, an error code, a clause. A search by meaning alone blurs these: it returns passages about similar things. Keyword search finds the passage that holds the exact term.
- Things said in other words. A question about how long records are kept should find a policy that speaks of retention periods and never uses the word "kept". Keyword search alone misses it. Search by meaning finds it.
A tool that makes you pick a mode makes you guess which kind your question is. Retiform's default is both, so the question is asked once.
What comes back
Passages, not whole documents. Retiform hands over passages and says where each came from. It does not write the answer, and a question the documents do not answer can still return passages on its topic.
- Each passage carries a citation: its document, the revision it belongs to and its position
in that revision.
read_passageturns a citation back into its text and says whether that revision is still the current one. assemble_contexttakes a question, searches by keyword and by meaning, and returns the best passages that fit a token budget, each with its citation. Your agent writes the answer from that context and cites the passages it used.searchreturns the matching records themselves when your agent wants to look before it reads. It takes a mode for the cases that need one. Left out, the mode is both.
Concepts describes documents, revisions, passages and citations in full.
An example
The sample corpus holds ten engineering records from a fictional company. One of its questions is:
Why were some passengers refunded twice during the August 2024 failover, and which rule did that break?
The answer is in two documents. The incident report says "the console's bulk refund action generated a new key on each click". A decision record written two months earlier holds the rule: "Every charge, capture and refund request carries an idempotency key". Your agent gets both passages, each with its citation, and you can open either one to check it.
From any agent
Your agent reaches search through the Model Context Protocol. The quickstart connects Claude Code or VS Code and asks its first question in a few minutes. Steps for other clients are in Connect other agents and are in beta.
What you can upload
Markdown, HTML, plain-text, Word and PDF files, through the console's Documents page, the kg
command or the REST API. A PDF is read page by page by a model, so it takes longer than other files.
A document can be searched by keyword as soon as it is stored, and by meaning once it is embedded,
usually within a minute.
After the search: keeping conclusions straight
What your agents conclude from the documents is kept too. A memory entry records who wrote it and cites the passages it rests on. A correction supersedes the earlier entry and keeps it readable. When a cited document changes, the entry is counted as citing a source that has changed, so a conclusion does not quietly outlive its evidence. Concepts covers memory.
What to know before you rely on it
- Size. The limits page gives the trial's limits. A search for several common words that must all appear can take seconds in a very large tenant; including a rarer word makes it faster.
- Where your text goes. Search by meaning needs embeddings, which OpenAI computes: it receives the text of your passages and of your searches. Security and trust says what it keeps and for how long.
- Hosted only. Retiform runs as a hosted service that keeps each team's data separate. There is no version to run on your own servers.
Try it
Retiform is in an invite-only trial, and the trial is free. Request a place on the home page.