Bucket Intelligence

Early access

Bucket Intelligence is in early access and is not generally available. Access is granted per account, so the feature does not appear in the dashboard unless your account has been enabled for it. If it is not available to you, contact support to ask about access.

It is free during early access. Pricing after general availability will be announced later.

Bucket Intelligence turns a storage bucket into a queryable knowledge base. Enable it on a bucket, and Fil One indexes the objects in that bucket so you can ask questions in natural language and get an answer grounded in your own files, with the source objects cited.

There are two parts: a background indexer that keeps each enabled bucket's index in sync, and a query endpoint that answers questions against it.

Enabling it on a bucket

1. Go to Bucket Intelligence in the left menu of the dashboard.
2. On the Buckets tab, click Index on the bucket you want, then confirm with Start indexing.
3. Wait for the first indexing pass to complete. Each bucket shows a status of Not indexed, Indexing, Failed, or Ready, and a Ready bucket also shows the number of files indexed, the index size, and when it last synced.

See Bucket Intelligence in the console for the full walkthrough of the dashboard, including asking questions and managing keys there.

Available in eu-west-1 and us-east-1.

How indexing works

Indexing is scheduled, not immediate

The indexer runs on a schedule, every 6 hours (03:00, 09:00, 15:00, and 21:00 UTC). It is not triggered by uploads. A file you upload now becomes searchable after the next scheduled pass, so allow up to 6 hours.

On each pass the indexer compares every object's ETag against what it indexed last time:

• New object — text is extracted, split into chunks, embedded, and added to the index.
• Changed ETag — the old chunks are removed and the object is re-indexed.
• Unchanged — skipped.

Deleted objects are reconciled out of the index on a full pass. Large buckets are checkpointed and resumed across runs, so the first pass over a big bucket may take several cycles to complete.

Text is split into chunks of roughly 1,000 characters with 200 characters of overlap.

Supported file types

Type	Extensions
PDF	.pdf
Plain text	.txt
Markdown	.md
HTML	.html, .htm
Word	.docx
PowerPoint	.pptx

Objects of any other type are ignored — they stay in your bucket untouched, they are just not indexed.

No OCR

PDFs are indexed by extracting their text layer. Scanned or image-only PDFs have no text layer, so they extract to nothing and are silently skipped — the file will not be searchable and no error is reported. Text-based PDFs work normally.

Models

Bucket Intelligence uses models that Fil One manages on your behalf, through Amazon Bedrock:

Role	Model
Embeddings	Amazon Titan Text Embeddings V2 (1,024 dimensions)
Answers	Anthropic Claude Opus 4.8

You cannot supply your own model or your own provider credentials, and you cannot choose a different embedding model. Bring-your-own-model support is planned but not available.

RAG API keys

To query a bucket from outside the dashboard you need a RAG API key — a bearer token, distinct from your S3 access keys.

What a RAG API key can and cannot do

A RAG API key authorizes only the query endpoint. It cannot read, write, list, or delete the contents of any bucket. It is also not a bring-your-own-LLM key — it authenticates you to Fil One, it does not carry credentials for a model provider.

Create one from the API Keys tab of the Bucket Intelligence page. You can scope a key to all buckets in your account, or to a specific list of buckets (up to 50). Bucket names are region-specific, so a scoped key names each bucket together with its region.

RAG API keys begin with sk_rag_. The token is shown once, at creation. Fil One stores only a hash of it, so it cannot be shown again — copy it immediately. The list view afterwards shows the key's name, a short prefix, its scope, when it was created, and when it was last used.

Keys do not expire

RAG API keys have no expiry and there is no rotation mechanism. A key stays valid until you delete it. Delete keys you are no longer using, and treat the token like any other long-lived secret — keep it in a secrets manager or an environment variable, never in source control.

Deleting a key takes effect immediately. Access also depends on the account that created the key: if that account loses early-access eligibility or its subscription lapses, its keys stop working.

Querying a bucket

From the dashboard

Open a bucket from the Buckets tab and use the Ask a question box in the bucket drawer. The answer appears with the objects it drew on listed as sources.

From the API

POST https://app.fil.one/api/buckets/{bucket}/query?region={region}

Authenticate with your RAG API key as a bearer token.

Field	Type	Notes
query	string	Required. The question to answer.
top_k	number	Optional. How many chunks to retrieve. Default 10, maximum 100.
model	string	Optional. Only the managed answer model is accepted.

region is a query parameter and defaults to eu-west-1. Add objectKey to restrict the search to a single object.

curl

curl -X POST \
  "https://app.fil.one/api/buckets/my-bucket/query?region=eu-west-1" \
  -H "Authorization: Bearer $FILONE_RAG_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "What is our data retention policy?"}'

Python

import os
import requests

response = requests.post(
    "https://app.fil.one/api/buckets/my-bucket/query",
    params={"region": "eu-west-1"},
    headers={"Authorization": f"Bearer {os.environ['FILONE_RAG_KEY']}"},
    json={"query": "What is our data retention policy?", "top_k": 10},
)
print(response.json()["answer"])

The response contains the grounded answer and the object keys it came from:

{
  "answer": "Records are retained for seven years, per the policy in compliance/retention.pdf.",
  "sources": ["compliance/retention.pdf", "policies/records-2025.docx"]
}

If nothing relevant is found, the answer says so and sources is empty. Requesting a bucket that does not exist, or one outside your key's scope, returns 404 in both cases.

Not available yet

Feature	Status
MCP endpoint / MCP server for Claude, Cursor, and other clients	Not built. Connecting AI clients is what the AI Agent Toolkit will cover.
Bring-your-own LLM keys or custom embedding models	Not implemented.
API key expiry and rotation	Not implemented. Keys are long-lived until deleted.
Upload-triggered or near-real-time indexing	Not implemented. Indexing is on the 6-hourly schedule described above.

Where your index is stored

Embeddings and the extracted text chunks are held in storage that Fil One manages, separate from your bucket. Your objects themselves are not moved or modified by indexing. The extracted chunk text is retained in the index so answers can quote from it, which is worth knowing if a bucket holds sensitive material — enable Bucket Intelligence deliberately, per bucket, rather than across everything.