Finding the right paragraph is not the same as chatting

TL;DR. Typing a procedure code, a name, or a sentence in quotes is a search for that paragraph. OrchAI Library cited chat looks with the words and the meaning together, keeps those exact terms, and looks once more if the first passages are thin — then it cites what it used.

The demo that felt finished

A knowledge lead sits through a demo called “search my files.” The question in the room is ordinary: where is the step for this job? Someone types the procedure’s code, or pastes one line from it in quotes.

A fluent answer comes back. It talks about the topic. It never opens the procedure that carries that code.

Picture a maintenance note filed as SOP-14, with the line “isolate the valve before you open the panel.” That picture is an illustration, so you can see the miss. It is not a score from a customer folder. A search that only follows meaning can land on a neighbouring note about safety and leave SOP-14 unopened. The code and the quoted line were the address.

Chatting about a topic and landing on the paragraph are different jobs. The useful one, for a practice lead, is the paragraph you can open and check.

Words and meaning, together

OrchAI Library is the memory of the suite — private Google for your team's documents, with every answer cited. The chat on a collection is the first place you ask. The product underneath is still the cited search.

On that cited chat, the usual question is searched both ways:

• By the words — the code, the name, the phrase you put in quotes.
• By the meaning — a plain restatement of what you asked.

Those wordings are prepared before the search runs. A conversational question can be restated once, in plainer terms, so the meaning side has something to hold. A comparison, or a question that needs two facts, can be split. The cap is three wordings. The code and the quoted phrase stay intact. A meaning-only pass is what drops them.

A collection can be set to words only, or to meaning only. That setting wins. When the collection has no such setting, cited chat uses both.

One more look, then the citation

The first set of passages is read for whether it is enough to answer. When it is thin, Library retrieves once more, with the wordings prepared again, and then it answers.

That is the whole extra step. Two retrieves at most. The answer cites the passage it used. You can open that passage and see whether SOP-14, or your own procedure, is actually there.

A longer mode that plans a search and keeps going is a separate switch. It stays off on this chat. Pilots still run on ask, then cite. This post is that path: words and meaning, one extra look, a citation.

The collection sits in a side panel

The collection itself is something you can manage in Library, from a side panel on the library list and on the collection page. If you administer it, you can create it, edit it, or delete it there, with a confirm step before a delete. Someone who only reads the collection does not see those buttons.

The panel is the collection. Getting the files in, and refreshing the index when a procedure changes, is the step beside it. A PDF dropped on a desk is not yet a cited paragraph.

Say this before anyone writes a purchase line

You can say today	Still needs a separate yes
Usual Library cited chat searches by words and by meaning together. A collection set to one mode keeps that mode.	“We measured how often your folder misses SOP-14.” The SOP-14 line above is an illustration.
A code, a name, or a quoted phrase is kept as words. Up to three wordings are prepared.	“Every wording becomes a new product mode.”
If the first passages are thin, Library looks once more, then answers and cites. Two retrieves at most.	“It keeps searching until the procedure appears.” The longer planning switch stays off.
People who administer the collection can create, edit, or delete it from a side panel. Readers do not get those buttons.	“Drop the PDF in the panel and it is indexed.” Loading files, and re-indexing when they change, is its own step.
The answer cites the passage. A pilot can run on that ask-and-cite path.	“OrchAI Eval has signed this search off.” A scored report is a separate product.

Where this sits

We wrote earlier about why a governed library cites the passage it recalls. A shared folder stores files. A library is there to recall a passage and name it.

We also wrote the recipe for a handbook: keep the clause whole, cite the section and the version, re-index when it changes. That post is how a policy is cut. This one is how a question finds the paragraph. Same library. Different job.

OrchAI Library holds that memory. OrchAI Agents is the front door people log into. OrchAI Eval is the quality gate when you need a scored report. OrchAI Portal is the full building, including who in the organisation may open what.

If the next conversation is “the demo searched our files and missed the procedure,” talk to us. Bring the code or the quoted line your team actually uses. We load that collection with you. The SOP-14 picture stays a picture.

Frequently asked questions

Why does a file search miss the procedure when I type its name? A search that only follows meaning can drift to a neighbouring note and drop the code, the name, or the sentence you put in quotes. Library cited chat keeps those words and also searches by meaning, unless that collection is set to one mode.

Will it keep searching until it finds the procedure? It prepares up to three wordings, retrieves, and may retrieve once more if the first passages are thin. Then it answers from what came back and cites the passage. It does not keep planning a longer search.

Can one collection stay meaning-only, or words-only? Yes. The usual cited chat uses both. A collection can be set to words only, or to meaning only, and that setting wins.

Does the second look mean the right paragraph is guaranteed? No score is promised here. The second look is a cap: one extra retrieve, then an answer with a citation. A separate evaluation sign-off is a different step.