Knowledge bases
A knowledge base is a set of documents Akili has indexed so that chat and agents can quote from them. Create one, add files or pages, wait for ready, and then pick it in the composer or link it to an agent.
Updated October 2026Library and knowledge2 of 3
Short answer
[Source N].Create a knowledge base
- Step 1
Open Knowledge
Choose Knowledge in the sidebar. The page lists your knowledge bases as cards, with a filter box and a usage meter that shows how many of your plan’s knowledge bases you have used. - Step 2
Name it
Choose New knowledge base, enter a name such as “Product documentation” and, if it helps, a description. Both can be edited later from the card. - Step 3
Expand the card
Click the card to open it. The expanded card is where documents, their status, the search box and the delete action live.
Add documents
Upload files
Drag files onto the expanded card or choose Upload. Plain text, Markdown, CSV, PDF, Word (docx) and HTML are indexed; HTML is stored as the extracted text. You can select several files at once. The size limit per file is the same per-category limit the Library applies, and uploads are capped at 60 per minute.
Every upload is mirrored into your Library and goes through the malware scan before indexing starts, so a document that is blocked by the scan never reaches the index.
Add a web page
Choose Ingest URL and paste a public https address. Akili fetches that one page, converts it to Markdown and indexes it as a document. It does not follow links, so add each page you need.
Add a Library file
From any file in the Library choose Add to knowledge base and pick the target. The dialog marks knowledge bases that already hold the file.
Connected SharePoint folders
A project with a connected SharePoint or OneDrive folder can index that folder’s files. Akili then maintains a knowledge base named after the project with a “SharePoint” suffix and keeps it in sync; it disappears when the folder is disconnected or indexing is turned off. See Projects.
Indexing status and reindex
Each document in the expanded card carries a status badge, and the card header counts how many are ready, processing or in error.
| Status | Meaning |
|---|---|
| pending | Queued. The document waits for the safety scan and a free worker. |
| processing | Text is being extracted, split into passages and embedded. |
| ready | Indexed. Chat and search can retrieve from it. |
| error | Indexing failed. URL documents show a Reindex button; re-upload a file to try again. |
Reindex appears on a URL document that ended in error. It fetches the page again and rebuilds its passages; the job retries three times with increasing delays before it reports an error again. For an uploaded file, delete the document and upload it again.
Search inside a knowledge base
The expanded card has a search box. Type a question or a phrase and Akili returns the closest passages, each with the document name, the passage number and its similarity score. Search is semantic: it matches on meaning through the embeddings, not on exact words, and returns up to 50 passages that score at least 0.6.
Use it to check what the index knows before you rely on it in chat, and to find which document a fact came from.
Use a knowledge base in chat
- Step 1
Pick the knowledge bases
In the composer, open the knowledge picker and select up to 20 knowledge bases. Each entry shows whether it is still indexing or ready. The selection is saved on the chat, so later messages keep using it. - Step 2
Ask the question
For every selected knowledge base Akili retrieves the five best passages that score at least 0.6, merges them into at most eight, and hands them to the model with the instruction to cite them. - Step 3
Read the grounding marker and the sources
A marker above the answer says grounded, no matches or indexing. The answer cites passages inline as[Source N], and a sources list under the answer shows the passages that were retrieved.
Use a knowledge base on an agent
In the agent builder, open the Knowledge section. You can upload files that belong to the agent alone, and you can link existing knowledge bases the agent should search. Linked knowledge bases apply to every conversation with the agent, including shared and published ones, so the builder reminds you that their content may be revealed in conversations. Files uploaded directly to an agent are kept in a hidden knowledge base that does not appear in your Knowledge list.
Plan limits
Three limits apply, all keyed to your plan: how many knowledge bases you can create (kb.max_count), how many documents each one can hold, and how much storage each one may use.
| Plan | Knowledge bases | Documents per base | Storage per base |
|---|---|---|---|
| Free | 3 | 50 | 100 MB |
| Pro | 10 | 500 | 1 GB |
| Max | 25 | Unlimited | 5 GB |
| Starter | 25 | 500 | 3 GB |
| Team | Unlimited | Unlimited | 10 GB |
| Business | Unlimited | Unlimited | 50 GB |
| Enterprise | Unlimited | Unlimited | Unlimited |
The usage meter on the Knowledge page tracks the first limit and turns amber at 80 percent. Creating a knowledge base past the limit opens the upgrade prompt. The document and storage limits are checked when you upload a file; the error names the limit you hit. See Plans and billing.
Retention and deletion
- Nothing expires on its own. Documents stay indexed until you delete them or the knowledge base. There is no retention timer and no automatic clean-up.
- Deleting a knowledge base is permanent. The confirmation names the number of documents; the documents and every passage derived from them are removed at once.
- Library and index are separate copies. Removing a file from the Library leaves the indexed document in place; removing a document from a knowledge base leaves the Library file in place.
- Account deletion removes everything. Knowledge bases are deleted with the account or organization that owns them, and a data-erasure request covers them too.
Good to know
- Projects can set defaults. A project with default knowledge bases applies them to every chat in the project unless that chat selects its own.
- Passages, not pages. Documents are split at blank lines into passages of roughly 512 characters. Documents with clear paragraphs and headings retrieve better than one long run of text.
- The mobile app has the same knowledge bases. You can create a knowledge base, upload documents and search from the Akili mobile app, with a per-base screen that the web app folds into the card.
- Advanced settings are API-only. The create endpoint accepts an embedding model, passage size and overlap; the web form sets the defaults.
Frequently asked questions
Can I share a knowledge base with my team?
Not directly. A knowledge base belongs to the person who created it. To give a team grounded answers, link the knowledge base to an agent and share the agent, or set it as a default knowledge base on a project that your colleagues are members of.
Does adding a URL crawl the whole site?
No. Ingest URL fetches that one page, converts it to Markdown and indexes it. Add each page you need, or upload an export of the site instead.
Why does the answer say no matches were found?
Retrieval only keeps passages whose similarity to your question is at least 0.6. Rephrase the question with the words the document uses, check that the documents are ready rather than processing, or search inside the knowledge base to see what it returns.
Which embedding model is used, and can I change it?
The default is OpenAI text-embedding-3-small, with passages of about 512 characters that overlap by 50. An organization can select another embedding model in Model Configuration, as long as it produces 1536-dimension vectors; the web form does not expose the setting per knowledge base.
Does deleting a file from the Library remove it from the knowledge base?
No. The knowledge base keeps its indexed copy until you delete the document there. Deleting the knowledge base itself removes every document and passage in it immediately.