Getting started
This procedure performs the first administration tasks, then confirms the installation with Smarty and a first SmartAgent.
Prerequisites
- Installation activated: Activate the installation.
- API key of the provider and references of the conversation and embedding models: Inference.
- Outbound HTTPS access from the server to the provider: Network.
- A non-confidential text document.
- Installation activated: Activate the installation.
- API key of the provider and references of the conversation and embedding models: Inference.
- Outbound HTTPS access from every Backend node to the provider: Network.
- A non-confidential text document.
Procedure
1. Add the providers and the models
To add a provider and its models, from the administration interface:
- Go to AI Factory, section Inference, page Instances.
- Open the Providers tab.
- Select the provider, click Configure and enter the API key.
- Open the models of the provider and enable the conversation reference, then the embedding one.
If embedding comes from another provider, add it the same way.
To choose the default embedding, before the first document upload:
- Go to AI Factory, section Capabilities, page RAG.
- In Default embedding model, click Change.
- Select the enabled embedding instance and save.
Expected result:
- the conversation and embedding instances appear and their connection test succeeds
- the Default embedding model banner shows the selected instance
2. Enable Smarty
Smarty, the general-purpose AI of the platform, is disabled at installation. To enable it, from the administration interface:
- Go to AI Factory, section Assistants, page Smarty.
- Open the configuration studio with the gear icon.
- In Models, add the conversation model to the Allowed models.
- Turn on In service in the top bar.
- Click Save.
Expected result: the Smarty page shows the Operational status.
3. Enable Alfred
Alfred, the built-in assistant that guides configuration, is disabled at installation. To enable it, from the administration interface:
- Go to AI Factory, section Assistants, page Alfred.
- Click Enable now.
- In Global model, choose the conversation model, then click Apply.
Expected result: no Alfred feature reports a missing model.
4. Set the default quota
By default, every user has 20 € per month for Smarty, the agents and AI Gateway. To adjust this budget, from the administration interface:
- Go to Governance, section Users, page Consumption.
- Click the Budget mensuel par utilisateur rule (monthly budget per user).
- In Limits, enter the monthly amount, then click Save.
Expected result: the Periods & limits column shows the new amount.
5. Test Smarty and a SmartAgent
This test confirms the installation end to end: models, assistants and document search.
To test Smarty, from the user interface:
- Open the home page and send a message to Smarty.
To create a first SmartAgent, from the administration interface:
- Go to AI Factory, section Assistants, page SmartAgents.
- Click Create an agent.
- In General, enter the Agent name, then click Next.
- In Models, choose the conversation model, then click Create.
To test the SmartAgent, from the user interface:
- Open SmartAgents, tab My agents, then the agent you created.
- Send it a message.
- Upload the document to a knowledge base, wait for its processing to finish, then ask a question whose answer is in it.
Expected result:
- Smarty and the SmartAgent answer
- the answer to the question cites the document
Note
If the conversation model is not offered when creating the agent, allow it in Governance, section Studio, page Models.
Special cases
Private inference farm for conversation
A private farm replaces the cloud provider for the conversation model. To add it, from the administration interface:
- Go to AI Factory, section Inference, page Instances, tab Providers.
- Add a farm and give it a unique name.
- Enter its full HTTPS URL, for example
https://inference.example.org/v1, and its API key. - Run the discovery of the served models.
- Select a known reference from the delivered catalog and enter its virtual input and output prices per million tokens, strictly positive.
- Import the model.
A private farm serves conversation. For embedding, choose a provider whose delivered catalog offers this capability.
The connection to the provider fails
| Cause | Check |
|---|---|
| TLS or proxy | DNS from the server, provider authority, Platform > Outbound network settings and allowed flows. |
| Access denied | Key, rights of the provider account and requested model. |
| Activation refused | Exact catalog reference, published capabilities and provider quota. |
OCR and images
These models are optional and separate from conversation and embedding. The catalog must offer a model of the right type, compatible with the installed package. For SaaS, only directly consumable model APIs are supported, without a serverless or dedicated endpoint to deploy.
Local OCR uses a detector and a recognizer on CPU, to be prepared from the catalog revisions before use: prerequisites. Testing a media instance counts towards consumption.
Check the entry points from the server
These commands query the API, the Frontend and the internal link between them. Replace smartagt.example.test with the configured name, and localhost with the server IP address if the API listens on it (Advanced YAML of Install the Backend). With the certificate generated by SmartAGT, -k is required. After Install your certificates, replace it with --cacert <authority of your certificates>.
curl --fail -k https://localhost:8000/api/v1/core/security/jwt/public-key
curl --fail -k https://smartagt.example.test:9000/ -o /dev/null
curl --fail -k https://smartagt.example.test:9000/api/v1/platform/internal-link
curl.exe --fail -k https://localhost:8000/api/v1/core/security/jwt/public-key
curl.exe --fail -k https://smartagt.example.test:9000/ -o NUL
curl.exe --fail -k https://smartagt.example.test:9000/api/v1/platform/internal-link
Expected result:
- the API returns its public key
- the Frontend answers with status
200 - the internal link reports
"available": true
Check the entry points of every node
These commands query the API of a Platform node, then the Frontend of a Frontend node and its internal link to the API. Run them again for every node, with its name. The API presents the certificate generated by SmartAGT by default, hence -k. The Frontend presents your certificate, verified by --cacert.
curl --fail -k https://<Platform node>:8000/api/v1/core/security/jwt/public-key
curl --fail --cacert <authority of your certificates> https://<Frontend node>:9000/ -o /dev/null
curl --fail --cacert <authority of your certificates> https://<Frontend node>:9000/api/v1/platform/internal-link
curl.exe --fail -k https://<Platform node>:8000/api/v1/core/security/jwt/public-key
curl.exe --fail --cacert <authority of your certificates> https://<Frontend node>:9000/ -o NUL
curl.exe --fail --cacert <authority of your certificates> https://<Frontend node>:9000/api/v1/platform/internal-link
Expected result:
- every API returns its public key
- every Frontend answers with status
200 - every internal link reports
"available": true