AI Search - Semantic – Papirfly Help Center

Discover semantic search

What is it?

Semantic search helps you find assets by describing what you are looking for in your own words.

Important: this is not a keyword search. Semantic search analyzes the visual content of the image to find what you describe.

For example: if you search for "a mountain at sunset", the system analyzes the images themselves (colors, shapes, composition) to find those with these visual characteristics. It does not look for these words in tags or metadata.

How does it work?

Images and search text are converted into vectors of 1024 numbers, in the same space. Search returns the images whose vector is closest to your query: you search by meaning, not by exact keywords.

Analyzed: shapes, colors, objects, contexts, moods. Optionally, selected metadata can also be vectorized for better relevance.
Not analyzed: exact keywords, business filters, technical references.
Model: Amazon Titan Multimodal Embeddings (amazon.titan-embed-image-v1), on Amazon Bedrock, AWS Paris region (eu-west-3).

When should I use it?

It works well for detailed visual descriptions:

"White SUV, side view, beige leather interior, photographed in studio"
"Navy blue silk dress on mannequin, zipper detail"
"18-carat gold ring with central sapphire, neutral gray background, studio lighting"
"White sandy beach, turquoise water, palm and coconut trees"
"Table set with white tablecloth, glass of red wine, gourmet plate"

→ The more detailed and visual the description, the better the results.

It does not work for:

product codes ("SKU564FRED");
specific dates ("published on March 15, 2024");
version numbers ("file 152A");
business statuses ("active", "archived").

→ For these cases, use the classic search with filters.

How do I activate it?

Semantic search is an optional feature, not enabled by default. It requires a specific configuration of your instance.

Contact your CSM (Customer Success Manager) or the support team.
Ask for semantic search to be activated.
Once enabled, the "Semantic search" button (AI icon) appears in your search bar.

How do I use it?

Open the search bar.
Click the "Semantic search" button (AI icon).
Describe what you are looking for in detail (colors, mood, context).
Refine with the classic filters if needed.

What configuration options are available?

Disable AI indexing for certain statuses or folders.
Limit semantic search to a selection of media.
Define the metadata fields to vectorize (4 to 5 maximum).

Frequently asked questions: architecture, security and billing

This section answers the questions most often asked by security and procurement teams.

Architecture and data flows

What is the architecture?

Keepeek is hosted by Claranet, in France. For semantic search, the Keepeek servers call a single external service: Amazon Bedrock, in the AWS Paris region (eu-west-3), through the endpoint bedrock-runtime.eu-west-3.amazonaws.com. Everything else (files, database, search index, vectors) stays at Claranet.

When is Amazon Bedrock called?

Three cases, all to the same endpoint:

Initial indexing: each existing image in your library.
Upload or replacement: each new or replaced image.
Search: each search performed in semantic mode.

What data is sent and received?

For an image: a rendition of the image in base64, 1200 x 1200 pixels maximum. Never the original file.
Optionally: the text content of the metadata selected for vectorization.
For a search: only the text entered in the search bar.
In return: a vector of 1024 numbers.

Where are the vectors stored?

At Claranet, in France, in a table and an Elasticsearch index dedicated to your instance. Amazon Bedrock keeps no data.

Does the data leave France?

No. Keepeek is hosted by Claranet in France, and Amazon Bedrock is called in the AWS Paris region. Semantic search data does not leave French territory.

Security

How are calls authenticated?

With a Keepeek technical credential (bearer token API key) sent in the HTTP headers, limited to invoking the model. Calls are made server to server: the key is held and operated by Keepeek and is never exposed on the client side.

How is data integrity guaranteed?

All calls go over HTTPS, with TLS 1.2 minimum enforced by AWS. TLS ensures the confidentiality and integrity of data in transit: any alteration invalidates the exchange. Calls are synchronous and direct, with no intermediary or queue. We do not add application-level signing beyond TLS; if your security officer would like an additional mechanism, we are happy to discuss it.

How is data separated between customers?

The platform is shared at the application level, and data is partitioned per customer:

a database schema dedicated to your instance;
an Elasticsearch index dedicated to your instance;
a vector table dedicated to your instance.

Some components can be deployed on dedicated servers depending on your configuration. On the Amazon Bedrock side, calls are stateless: nothing is kept between calls. Model providers have no access to customer accounts, logs or data.

Is my data used to train models?

No. Neither AWS nor third-party model providers use Bedrock inputs or outputs to train any model (Amazon Nova, Amazon Titan or third-party models). Your data and images therefore cannot be reused to train models intended for other customers. Reference: Amazon Bedrock FAQ.

What contractual framework applies (GDPR)?

The AWS Global DPA applies automatically: it is incorporated into the AWS Service Terms. The corresponding contractual commitments can be provided on request.

Billing

How is semantic search billed?

It is included: no billing per image, per process or per token. There is therefore no consumption cost to estimate. To activate it, contact your CSM.