Agency File Management in 2026 — Why Google Drive Alone Fails and What Kobin Vault Does Instead

What does the agency file chaos problem actually look like?

If you run a 3–15 person agency, your file situation has a specific shape. The client brief arrived via email. You moved it to a Google Drive folder. The designer put the first draft in a different Google Drive folder under a name that made sense at 11pm on a deadline night. The revised version is in Dropbox. The approved version was screen-shared in a Zoom call and never formally saved. The final version you actually delivered lives in the client's own inbox as an email attachment — and you have no copy.

This is not laziness. It is the predictable output of a tool stack that was never designed to handle agency delivery workflows. Google Drive is file storage. Notion is flexible databases. Slack is messaging. None of them have any concept of a project-scoped delivery pipeline — where some files belong to the team, some belong to the client, and some are handoffs between the two.

The specific failure modes show up in every agency at every size. Here are the seven most common ones — and what Kobin Vault does instead:

How is Kobin Vault structured? The three-folder architecture.

Kobin Vault uses a three-folder hierarchy per project, created automatically when you create a project. Every project gets exactly three folders: Internal Documents, Client Uploads, and Deliverables. These are not arbitrary labels — they map to access control rules enforced at the database level.

• Internal Documents is visible only to your team (founder + team members). Clients cannot see it, cannot navigate to it, cannot accidentally stumble upon it. This is where you put your internal briefs, strategy documents, SOPs, call notes, and anything else that is working context rather than client-facing output.
• Client Uploads is visible to both team and clients. Clients can upload files directly into this folder through their portal — brand guidelines, approved photography, legal documents, reference materials. Team members can see everything clients upload. This eliminates the “email me your brand kit” workflow that scatters assets across inboxes.
• Deliverables is where the work goes when it is done. Team members upload to this folder; clients can view and download but cannot modify. When a task with a deliverable requirement is completed, the upload automatically routes here. The client portal reflects this immediately — no manual notification, no shared Drive link, no email attachment.

These folders are backed by Google Drive using the drive.file scope. This means Kobin only accesses files it creates — it never reads your existing Drive content. Your files stay in your own Google account. If you cancel Kobin tomorrow, your files are still in Drive, exactly where they were.

How does AI auto-labeling work on upload?

When you upload a file to Kobin Vault with AI labeling enabled (the default), the system extracts the full text from the document before it ever hits the server — and asks an AI to label it for you. The result is that you never manually type a title, description, or document type for a file you upload. The AI does it from the actual content, not the filename.

The extraction pipeline runs client-side for most formats. DOCX files are parsed with Mammoth, which converts Word formatting to plain text. XLSX and CSV files are parsed with the xlsx library, producing a sheet-to-CSV representation that preserves all column values. Code files (JavaScript, TypeScript, Python, and 25+ others) and plain text files are read directly. PDFs are handled server-side after upload via pdf-parse, which reliably extracts text from most PDF formats.

Up to 3,000 characters of the extracted content are sent to a fast AI model (Groq's Llama 3.3 running via the HuggingFace router) with a strict JSON-only system prompt. The model returns four fields: title (a human-readable name derived from the actual content), description (one sentence describing exactly what the document is), document_type (one of 14 categories: Content, Deliverable, Report, Contract, Brief, Design Asset, Spreadsheet, Presentation, Reference, Proposal, SOP, Note, Code, Other), and tags (2–4 lowercase tags for additional categorization).

The form pre-fills automatically while the analysis runs. A small “Labeled” indicator appears when it is complete. You can review the AI's suggestion, edit any field if needed, and click Add to Vault. If AI labeling is turned off, the full manual form is shown instead.

How does semantic search work in Kobin Vault?

Every file stored in Kobin Vault is automatically vector-embedded and indexed for semantic search — meaning you can search by concept and meaning, not just filename or keyword. Press ⌘K from anywhere in the workspace to open the search overlay and type a natural-language query. Results appear in under 2 seconds.

The embedding pipeline runs in two layers. First, the full document — title, description, extracted text, and any note content — is embedded as a single vector using the BAAI/bge-small-en-v1.5 model (a 384-dimension retrieval model fine-tuned for semantic similarity). This item-level embedding powers fast overview search.

Second, if the document has substantial text content (more than 200 characters), the full text is chunked into 800-character segments with 160-character overlap — a 20% overlap that preserves semantic continuity across chunk boundaries. Each chunk is embedded separately and stored in a vault_chunks pgvector table. When you search, Kobin runs a chunk-level similarity query first (higher precision), then deduplicates by document, keeping the best-matching chunk per file. The final results pass through a Max Marginal Relevance algorithm that balances relevance to your query against diversity — so you get results from different projects and folders rather than five chunks from the same document.

Hybrid mode (the default in the ⌘K overlay) runs semantic search and keyword search in parallel, merges and deduplicates the results, and ranks by semantic score. This means a query for “Reelix payment terms” will find the contract even if you never typed “payment terms” as a tag — because the vector knows what the document says.

What file types can you view and edit inside Kobin Vault?

Kobin Vault includes seven different native viewers and editors, covering every file type an agency regularly handles. You never need to download a file and open it in a separate application to review or edit it. Everything opens inside the workspace.

The Monaco Code Editor

For code files, Vault uses Monaco — the same editor that powers Visual Studio Code. When you open a TypeScript file, a Python script, a SQL migration, or any of the 30+ supported languages, you get full syntax highlighting in a custom Kobin Dark or Light theme. You can toggle between View and Edit mode, make changes directly in the browser, and save with Cmd+S. The editor detects the language automatically from the file extension. A revert button appears when there are unsaved changes, letting you roll back to the last saved version.

The Monaco AI Writer panel (accessible via the Kobin AI button in the toolbar) gives you a code-specific AI assistant. You can type instructions like “add TypeScript types to all functions,” “add error handling and try/catch,” or “refactor for readability” — or choose from pre-built suggestions. The AI generates code that you can either append to the existing file or replace it entirely. When you save, the new content is re-embedded and the search index updates.

The TipTap Document Editor

Notes, Markdown files, plain text documents, and items labeled as Brief, SOP, Reference, or Proposal all open in a full TipTap rich-text editor. The editing experience is similar to Notion or Linear's document view — you get a large title field at the top, a clean editing canvas, and a floating toolbar with formatting options. The toolbar supports Bold, Italic, Underline, Strikethrough, H1–H3 headings, bullet lists, ordered lists, todo checkboxes, inline code, blockquote, and horizontal rule.

Markdown files are automatically converted to HTML on open for proper rendering — a heading written as ## Section displays as a formatted H2. The AI Writer panel connects to the RAG engine: when you ask it to expand a section or draft new content, it searches your entire vault for relevant context first, cites the sources it used, and streams the response in real time.

The Native Spreadsheet Editor

Excel files (XLSX, XLS), CSV, and TSV files open in a native spreadsheet viewer built with the xlsx library. All sheet tabs are displayed. The view shows paginated rows (50 per page) with column headers. In Edit mode, every cell becomes an inline input — click a cell, type your change, Tab to the next cell, Enter to move down, Esc to confirm. Header cells are also editable. When you save, the active sheet is exported as a CSV and stored in extracted_text, making it immediately searchable and AI-readable.

The AI Writer — Kobin Vault as a NotebookLM for your agency

Kobin Vault's AI capabilities go beyond search. The Vault AI Writer can read the content of your documents, synthesize information across multiple files, and write or edit content that is grounded in your actual project knowledge — not hallucinated from general training data.

When you open the AI Writer panel in any document, you can ask questions or give writing instructions. Before the model generates a single word, it runs a semantic search across your entire vault to retrieve the most relevant documents. Up to six vault items are injected into the AI context as a structured knowledge brief — titled, typed, and attributed. The model is explicitly instructed to cite the source document when it uses information from it, and to say so if no relevant vault context exists for the question.

This is the same architecture that makes products like Google NotebookLM useful — grounding AI responses in your own documents rather than the model's training data. The difference is that Kobin's vault AI is not a standalone product. It is embedded directly in the editor where you are working, and the knowledge base is the same vault you use every day for client delivery. When the AI writer cites “Reelix Brand Brief — Brief (91% match)” as its source, it is citing a specific file your team uploaded — not a generic web result.

The AI Writer also works in the global command bar (⌘K). You can ask questions like “what does our standard SLA say about response time?” or “find the scope of work for the Acme Corp project” and the AI will retrieve the answer from your vault files. This is how the Vault transforms from a file storage system into a genuine knowledge base for your agency.

What does the Vault right panel include?

Every open vault item has a collapsible right panel with four tabs: Context, Approval, Comments, and Activity. Together these tabs give you the full operational picture of a file — not just its content.

• Context shows the AI memory status (whether the file has been vectorised and indexed), a list of related vault items by semantic similarity (the four most similar files across all your projects, with percentage scores), and quick action buttons for drafting follow-up emails, creating delivery tasks, or finding similar files across projects.
• Approval is available for files in a Deliverables folder. It shows the current approval status (Awaiting, Approved, or Changes Requested), a note field for the approval reviewer, and one-tap approve / request-changes buttons. For PDF and image files, an inline file preview is shown in the approval tab so reviewers do not need to download the file to approve it. A signed 1-hour URL is generated server-side for the preview.
• Comments is a threaded discussion attached to the specific vault item. Team members can comment on deliverables, ask questions about briefs, or leave revision notes directly on the file — keeping the conversation co-located with the content rather than in a separate Slack channel.
• Activity shows a full timeline of everything that has happened to this file: when it was added to the vault, when it was edited, when it was vectorised by the AI, when it received comments, and who did what. The activity log is automatically populated — no manual logging required.

How does Kobin Vault change the client file delivery workflow?

The traditional agency client delivery workflow has six manual steps: complete the work, upload to Drive, copy the shareable link, open Slack (or email), paste the link with a message, and wait to confirm the client received it. Each step is a potential failure point. Drive links go stale. Slack messages get buried. Clients lose the email.

Kobin Vault eliminates all six steps. Here is the actual workflow:

1. A team member marks a task as complete. If the task has a deliverable requirement, a completion modal opens — they cannot close it without uploading the file.
2. The file is uploaded to Supabase Storage via a private vault bucket. A storage path is generated using founderId/folderId/timestamp_filename.
3. The file appears immediately in the Deliverables folder of that project vault.
4. The client is already logged into their portal (or next time they log in). They see the file in their Deliverables folder without any notification, link, or email.
5. The client can open the PDF in the inline viewer, review it, and mark it Approved or Request Changes — directly in their portal.
6. The Activity log on the vault item records the upload, the view, and the approval decision.

This is not an incremental improvement on the Drive + Slack workflow. It is a structurally different architecture that removes the human handoff entirely. The file goes from “completed task” to “visible to client” without any manual step.

Kobin Vault vs Google Drive vs Notion for agency file management

How does Kobin Vault handle security and access control?

File security in agencies has a specific risk profile: the risk is not external hackers. It is internal exposure — a client accidentally seeing another client's files, or an internal strategy document being shared with the wrong person.

Kobin Vault addresses this structurally, not procedurally. The access control is embedded in the folder type — not in manual permission toggles that someone has to remember to set on every upload. Internal Documents are inaccessible to clients at the database query level. Client Uploads are visible to the specific client who is associated with that project — not to clients on other projects. Deliverables are accessible to that client only.

Files are stored in Supabase Storage's private vault bucket. They are never directly accessible via public URL. When a viewer needs to load a file — for example, when a client opens a PDF in their portal — a signed URL is generated server-side with a 1-hour expiry. The signed URL generation checks that the requesting user has access to the item before returning a URL. Even if a URL were somehow shared with the wrong person, it would expire within the hour.

For team members, the access model mirrors the founder's project access — a team member can only see vault items in projects they have been added to. For clients, the access is even more constrained: they only see the Client Uploads and Deliverables folders for their specific project. The Internal Documents folder does not appear in their portal at all.

What does it actually cost to replace Google Drive and Notion with Kobin Vault?

Subscription costs from published pricing pages, April 2026. Manual overhead calculated at $75/hr blended rate, 3 hours/week for a 5-person agency. Actual savings vary by team size and current tool stack.

The subscription argument understates the real saving. The more significant number is the recovery of the 3+ hours per week that the average agency loses to manual file management — finding files, sharing Drive links, chasing client approvals, updating version names, and fixing access permissions. At a blended $75/hr rate for a 5-person team, that is nearly $1,000 per month in invisible overhead. Vault's structured delivery pipeline, semantic search, and automatic client portal integration eliminate the majority of those manual steps.

Frequently Asked Questions