Collections

A collection is a searchable set of documents, embedded by one model and loaded by its pipelines. Agents and assistants answer from collections. Builders create them, fill them and test them here.

Who can open it

Role	What they can do
Admin	Create, load, ask, search, change the graph setting and delete collections.
Builder	The same as an Admin.
Auditor	Open the list and every collection page and read them. Cannot add documents, ask, search, change or delete.
Reviewer	Cannot open it.
End user	Cannot open it.
Platform admin	Everything an Admin can, for any tenant.

An Auditor who opens a collection’s Graph and search tab sees the graph and this line in place of the search: “Your role can read the collection, not query it.” Asking and searching spend the tenant’s tokens.

What you see

The collections list

Build › Knowledge opens on the Collections tab. It lists every collection in the tenant, across all of its vector databases, with a search box and, for Admins and Builders, Create collection.

Column	What it shows
Collection	The collection’s name. Opens its page.
Vector database	The database that holds it. Opens the list filtered to that database.
Chunks	How many chunks it holds. A chunk is one stored passage of a document.
Loaded by	The pipelines that load it and how each one’s last run went, No pipeline when none does, or Could not be read.
Actions	Create pipeline and Delete collection, for Admins and Builders.

When the list is filtered to one database, a line says “Showing the collections in name” with Show all collections beside it.

Empty and error states:

• No knowledge collections yet: the tenant has no vector database. Create vector database opens that wizard. See Vector databases.
• No collections yet, “A collection can be created once one of this tenant’s vector databases is ready”: databases exist, but none has finished provisioning. View vector databases switches tabs.
• No collections yet (or No collections in name yet when filtered): a ready database exists. Create collection starts the wizard.

A collection’s page

The header shows the collection’s name and one line: the database it is in, the embedding model and its dimensions, and the chunk count. Admins and Builders also see Add documents, Create pipeline and Delete collection.

The page has five tabs:

Tab	What it holds
Overview	Its facts (Vector database, Embedding model, Dimension, Chunks, Created), a summary of its graph with Rebuild graph, and Recent activity: the latest pipeline runs into it, each with its documents and chunks.
Documents	Add documents (Admins and Builders), then every document its pipelines have loaded, newest first, with Filename, Loaded through, Status, Chunks, Size and Added.
Pipelines and runs (n)	Each pipeline that loads it, with Pipeline name, Source, Parser and Last run, and Create pipeline for this collection.
Graph and search	The collection graph, Ask and Raw vector search. See The collection graph.
Settings	Set when it was created (Name, Embedding model, Dimension, Distance), the People in the graph switch, and Delete collection.

Text you paste into a collection is stored without a pipeline: it counts among the chunks but is not listed on the Documents tab.

If the collection was deleted, or belongs to another tenant, the page says This collection is not here.

What you can do

Create a collection

You need a vector database whose status is Ready, and an embedding model in the model catalog.

1. Open Build › Knowledge. On the Collections tab, choose Create collection.
2. Collection step:
   1. Vector database: the database that holds it. The wizard picks it for you when the list is filtered to one database or only one is ready.
   2. Collection name: lowercase letters, digits and underscores, starting with a letter, at most 48 characters. When a name is not allowed or already used in that database, the wizard offers a corrected one: choose Use name.
   3. Embedding model: the model that embeds every document in it. If the catalog has none, the step says The catalog has no embedding model.
3. Ingestion pipeline step: turn on Create an ingestion pipeline now to create its first pipeline in the same pass, then enter the Pipeline name, the Parser, Chunk size (chars) and Chunk overlap (chars). Leave it off to load documents later.
4. Source and schedule (only with a pipeline): where the documents come from and when the pipeline runs. See Pipelines and connectors.
5. Documents (optional): drop files to load once the collection exists.
6. Review: check what will be created, then choose Create collection, or Create collection and pipeline.

Olympusx AI opens the new collection’s page, on its Documents tab when you dropped files.

When creating the collection works and creating its pipeline does not, the wizard says “name was created, but its pipeline was not” with the reason. Go back, correct the pipeline, and choose Create the pipeline: the collection is not created again.

Caution

The embedding model, the dimension and the distance are fixed when the collection is created. Every vector in a collection comes from one model. To embed with another model, create a new collection and load the documents into it.

Olympusx AI measures the dimension by sending the model one test embedding when you create the collection. A model served by your tenant must therefore be deployed and ready, unless the catalog already records its dimension; otherwise the create is refused with “Deploy the model, wait until its deployment is ready, and create the collection again.” A cloud provider model must be granted to the tenant and must answer. The test embedding is metered like any other call.

Add files

1. Open the collection and choose Add documents, or open its Documents tab.
2. Drag files onto Drag and drop files here or click to choose them, or click to pick them.
3. Olympusx AI sends the files one after another and shows “Sending file: n of total”. Each row ends loaded, with its chunk count, or Not loaded with the reason.
4. To resend the files that failed, choose Try the failed files again (n). Clear the list empties it.

The first file you add creates an upload pipeline of the collection’s own. Every later file goes through it: parsed, cut into chunks and embedded with the collection’s model.

Which file types are read depends on the parser:

Parser	File types
Local (text, Markdown, HTML)	.txt, .text, .md, .markdown, .html, .htm, .csv, .json, .log
The document-converter parser (its label reads “PDF, DOCX, scans”)	All of the above, plus .pdf, .docx, .pptx, .xlsx, .adoc, .asciidoc, .png, .jpg, .jpeg, .tif, .tiff, .bmp, .webp

The document-converter parser is offered only when your installation includes the document converter. Without it, PDF, Office documents and images are refused with “This type needs the … parser, and this installation does not have it.”

Paste text

1. On the Documents tab, choose Paste text.
2. Type or paste text, one document per line.
3. Choose Add the text. Olympusx AI reports “Documents added: n. Tokens embedded: n.”

Each line is stored as written: it is embedded as one chunk, without parsing or chunking. Use this for short facts; use files or a pipeline for anything longer.

Ask a question

Open the Graph and search tab. Admins and Builders see two tools beside the graph.

1. Under Ask, choose the Answer model: a chat model deployed for your tenant.
2. Type a question and choose Ask.
3. Olympusx AI retrieves passages by meaning and by keyword together, reorders them with the platform’s reranker model when one is configured, and has the answer model write an answer that cites them as [1], [2]…
4. Under Sources, the cited passages are listed first, each tagged Cited with its relevance. Show passages retrieved but not cited (n) shows the rest.

If no chat model is deployed for the tenant, the tool says No answer model deployed. Deploy one from Models.

Search without an answer

1. Under Raw vector search, type what to look for.
2. Choose Search. Olympusx AI returns the 5 chunks closest in meaning, each with its source and score, and lights them in the graph.

Load from another source

Choose Create pipeline, or Create pipeline for this collection on the Pipelines and runs tab. The pipeline wizard opens with this collection already chosen. See Pipelines and connectors.

Delete a collection

1. Open the collection, then Delete collection (from the header or the Settings tab), or use Delete collection in the list’s Actions menu.
2. Type the collection’s name to confirm, then choose Delete.

Its documents and their embeddings are deleted and cannot come back. The pipelines that load it are deleted with it, and agents and assistants that answer from it lose it as a source.

Limits

What	Limit
Collection name	Starts with a lowercase letter; lowercase letters, digits and underscores; at most 48 characters; unique in its vector database
Embedding model	One per collection, fixed at creation
Distance	Cosine, fixed at creation
One uploaded file	At most 16 MiB
Files in one drop	Sent one after another, each through the collection’s upload pipeline
One document	At most 2,000 chunks; a larger one fails
Pasted text, one request	At most 1 MiB; each line is one chunk
Ask	6 passages go to the answer model
Raw vector search	Returns 5 chunks
Budget	Loading, asking and searching are refused once the tenant’s budget for the period is spent

Related

• Knowledge
• Vector databases
• Pipelines and connectors
• The collection graph
• Create and edit an agent
• Keys and budgets
• Knowledge and grounding