Files & knowledge

SupaNet has a Files area where you and your team can store documents. PDFs get special treatment: once uploaded, the assistant can read them and answer questions from their contents.

Uploading files

Drop files into the Files page (or attach them in chat). Each file is stored privately under your account.

You can view your files in list view (default) or grid view with thumbnails — toggle between them with the buttons near the Upload button. Your preference is saved and persists across sessions.

Sharing files

Select one or more files and choose a share mode from the toolbar:

• Public — publishes the file to a permanent public link. The file is copied to a separate public bucket, so you can bake the URL into a public artifact or static HTML page. The original private file is untouched. Once published, a "Public" badge on the file lets you copy the link or unpublish it (which removes the public copy).
• 1 hour / 1 day / 1 week — creates a temporary signed link that expires after the chosen window. These are useful for sharing with external parties for a limited time. Signing still only marks the file as "link shared" — the private original stays in the private bucket.

A results modal shows all the generated links so you can copy them individually or all at once. For a single file, the link is automatically copied to your clipboard.

Adding titles and descriptions

When you upload a file, SupaNet assigns it a filename. You can add a title and description to help organize and find files later. Click on any file's name to edit:

• Title — a human-readable name that defaults to the filename if not set
• Description — optional long-form notes or context about the file

These fields are searchable and help teammates understand what each file contains at a glance. They are also used when linking files into artifacts (see below).

From files to compiled knowledge

SupaNet transforms raw documents into maintained understanding.

Instead of searching PDFs from scratch on every question — which means the assistant re-interprets everything, context repeats, and contradictions hide — SupaNet can compile your files, links, notes, and messages into a knowledge layer. Compiled pages are durable reference prose (concepts, decisions, processes, people, projects) with evidence cited inline. They accumulate. They remember.

Two approaches: search, or compile

Search works when you need the exact wording or you're still gathering information. You upload a PDF, SupaNet indexes it, and the assistant pulls in relevant passages when you ask a question.

Compile works when you want to maintain living reference material. The compiler interprets your files, links, notes, and messages into a structured knowledge layer, detects contradictions, and flags when information gets stale — so answers are grounded in stable, cited understanding instead of re-read raw documents.

You can use either, both, or neither.

How search works

When you upload a PDF, SupaNet does the following in the background:

1. Pulls the text out of the PDF.
2. Breaks it into small chunks.
3. Builds a searchable index using hybrid retrieval — both semantic search (understanding meaning) and keyword search (matching exact terms).

The assistant can then search across your PDFs and pull in relevant passages when you ask a question. It tells you which document each answer came from and when it was last updated, so you can trust and assess freshness.

How hybrid search works

The knowledge base runs a single hybrid search on every query, combining two signals:

• Semantic search finds passages that are conceptually related to your question, even if the exact words don't match. For example, searching for "refund policy" will find a passage that says "money-back guarantee".
• Keyword search finds exact matches for terms, IDs, and names that semantic search might miss. For example, searching for "SKU-12345" will find that exact product code.

These two signals are fused using a ranking algorithm that surfaces chunks appearing in both searches at the top. This hybrid approach is more accurate than either method alone — it catches paraphrases semantic search misses and exact terms keyword search wouldn't find.

You do not have to do anything to trigger this. Upload the PDF and, a short time later, it is searchable.

Compiling knowledge

Knowledge compilation turns a collection into a maintained reference layer. To start compiling:

1. Go to Knowledge (sidebar, under Assets).
2. Choose a collection in the Policy tab.
3. Set the collection's compilation settings — especially Autonomy, which controls how freely the compiler can write:
   • Suggest — propose everything to you for review
   • Guarded (default) — create and add freely, ask before rewriting
   • Auto — create, add, and rewrite freely (ask only before wholesale replacement)
4. Click Compile now, or schedule a nightly pass via Automation.

The compiler then:

• Gathers new files, links, notes, and messages in that collection
• Extracts structured understanding (concepts, decisions, people, processes, etc.)
• Updates your compiled pages or creates new ones
• Flags contradictions and stale entries for human review
• Emits a change brief so you can see what changed

The Knowledge dashboard

Knowledge (Assets) has four tabs:

• Review — conflicts and held updates. Leads on purpose because contradictions matter.
• Compiled — your maintained pages, grouped by kind (concept, decision, process, person, project, terminology, principle, question, profile).
• Briefs — what each pass changed, with a live checklist while one runs.
• Policy — edit the trust boundary per collection, and trigger compilation.

Trust and contradictions

Compilation is not unrestricted. The trust boundary controls:

• What gets compiled — files, links, messages, artifacts, to-dos, notes
• What page kinds — concepts, decisions, people, processes, decisions, etc.
• How much autonomy — from "suggest everything" to "rewrite freely"
• Safety guards — page kind, labels, or title (e.g. mark "financial commitments" as always-review)

When new evidence contradicts existing knowledge, the compiler never decides who is right. It flags the contradiction, marks the page disputed, and asks you. Options: use the new source, keep the old one, or dismiss.

This is why: a bad source in a one-off search is isolated. A bad source folded into compiled knowledge spreads into every answer built on it. So contradictions are caught explicitly, not swept under the rug.

Shared by default, private when you want

By default, the knowledge from a PDF is shared with your workspace - any teammate's chat can find and use it. That is usually what you want: one person uploads the handbook, everyone can ask about it.

If a document is sensitive, the owner can flip it to "Only me" so only they can use its contents.

Two important details:

• This sharing is about the searchable text, not the raw file. The original PDF stays private to whoever uploaded it. Only the extracted passages are shared.
• Only the owner can change whether a document is shared or private.

Adding content from the assistant

The assistant can save files directly to your Files area. This is especially useful for binary output it generates — like a base64 image from an image model — which it can now persist as a real, shareable file instead of just showing you the raw bytes.

The assistant can:

• Create files from base64 (images, PDFs) or plain text
• List and fetch files it previously created
• Organize files into collections
• Tag files with metadata (e.g. source provenance)

If it fetches an article or you ask it to turn a transcript into notes, it can save that content so the whole team (or just you, if you prefer) can search and learn from it later. PDFs are automatically indexed into the knowledge base.

Adding files to collections

Files can be added to collections — named groups of related content you can chat with. On the Files page, select the files you want and add them to a collection (create a new one or pick an existing one). Then in Chat, use the 📚 picker to select that collection; the assistant will have access to all the files, artifacts, and to-dos in the collection when answering your questions.

This is useful for grouping related documents around a topic, project, or process. For example, you might create a "Customer Handbook" collection with the main handbook PDF and several supporting documents, then chat with that collection to answer handbook-related questions all at once.

Putting it together

The practical workflow looks like this: someone uploads the policies, the contracts, the spec sheets. The assistant can pull articles and notes from the web and add them to the knowledge base too. From then on, anyone on the team can open the chat and ask "what does our refund policy say?" and get an answer grounded in the actual document, with a citation pointing back to it. You can also group related files into a collection and chat with them all at once as a focused context.