Immersive mode puts an assistant beside the form your team is filling. It checks every field against your rules and your documents, drafts fixes people approve, and holds the finish until blocking issues are resolved.
Flag the problem on the field. Draft the fix. Hold the finish. Keep the trail.
A sample invoice, drawn by the SDK's own components. Click Approve and Suggest fix.Bring your own model key
- Anthropic
- OpenAI
- OpenRouter
Immersive mode
The assistant that works beside your team, not behind a chat box
Chat answers questions. Immersive mode watches the work. While someone has an invoice, a claim or a contract open, Baseera sees its fields and your checks, and speaks up on the field that is wrong.
Flags on the field
Rules in your app and AI checks that read your documents raise blocking, warning or info flags on the exact field at fault.
Drafts people approve
The assistant proposes a value, such as a tax ID from the customer record. It reaches the form only when the person approves it in the field.
A finish that holds
Blocking issues disable the finish. Someone can override with a written reason, and your server verifies a single-use receipt before it acts.
A trail for every record
Every flag, fix, approval and finish is kept per record, shown as a timeline and sent to your server as a signed webhook.
I want to add Baseera immersive mode to the invoice screen in this app. Read https://baseera.dev/docs/llms-full.txt first and follow it; do not guess APIs. 1. Find the invoice form and list its fields (customer, tax ID, total, and any others). 2. Install the Baseera packages the docs name for this framework, and wrap the app in the provider. The publishable key and base URL come from environment variables. Ask me for the values and never write a secret into the code. 3. Add the user-token route on the server, so the browser only ever holds the publishable key. 4. Declare the record's fields and checks for immersive mode: a blocking check when the tax ID is missing, with a draft fix taken from the customer record, and a warning when the total looks wrong. 5. Make the Send button the guarded finish, and verify the receipt on the server before the invoice is sent. 6. Run the type-check and the app, show me the screen with a blocked invoice, and tell me what you changed. Ask before anything that is not in the docs.
Hand the setup to your coding agent
Paste this prompt into the agent you already use. It reads our docs, finds your screen, and wires up the fields, the checks and the guarded finish.
You stay in control: the agent asks for your keys, and the finish is enforced on your server as well.
Everything your assistant needs to answer, act and prove it
Baseera helps your team build an assistant that answers from your own knowledge, acts inside your app, and proves it works, all in one secure workspace.
Search and retrieval
Discover
From your documents to an answer you can check
The assistant does not guess. It searches your documents first, cites what it found, and says so when it finds nothing.
Actions with approval
Act
It acts inside your app, and the user decides
Define the tools the assistant may call. Any tool that changes data asks for explicit approval, and can be undone after it runs.
- Approve or reject, with an optional reason
- Undo after it runs
- A record of every step
Open a return for headphones, order 0421
- Order
- #0421
- Reason
- Withdrawal within 14 days
- Refund
- To the original payment method
Testing and review
Measure
Measure the assistant before customers meet it
Every change to instructions, model or knowledge is tested before it ships, and watched after.
Test suites
Cases with written expectations: whether a tool was called, the reply text, the language, steps, cost and time.
Model judge
A model grades the answers a rule cannot, against criteria you write.
Online evaluation
Samples of real conversations are graded automatically, so you see when quality drops.
Human review
Reviewers mark an answer correct, wrong or unsafe, and write the right answer.
Compare versions
Run the same suite on two versions and see what improved and what regressed, before you publish.
Where trust meets measurable quality
The product is built around a few rules that do not bend. These are the numbers behind them.
- 4Model providers, each with your own key
- 2Languages, Arabic first
- 5Ways to evaluate an answer
- 0Actions taken without the user's approval
Private. Secure. Tested.
Built secure from the first table
We assume mistakes happen, so protection lives in the database and the server, not only in the interface.
Isolation between organizations
Every table has row-level security, and the build fails when a table has none.
Your own keys
Provider keys are encrypted in a vault, never returned by any API after saving, and never written to logs.
Short-lived tokens
User tokens are signed with EdDSA only, other algorithms are refused, and a revoked key stops working at once.
Knowledge that cannot leak
Search filters inside SQL by organization, tenant and access rules, so a passage never reaches someone who may not read it.
Spend and rate limits
A monthly spending cap and request limits keep a wrong experiment from becoming a bill.
Signed webhooks
Events carry an HMAC signature and a timestamp, and stale messages are rejected.
Arabic first, not translated later
Right-to-left layout, numerals, dates and the wording of replies are built for Arabic, then mirrored for English. The same components work in both directions.
أنت: اشتريت سماعة قبل 10 أيام وأريد إرجاعها. هل ما زال ذلك ممكنًا؟
بصيرة: نعم، يمكنك إرجاعها خلال 14 يومًا من الاستلام بشرط أن تكون بحالتها الأصلية 1. رسوم الشحن العكسي على المتجر إذا كان العيب من المنتج 2.
- سياسة الاسترجاعالإرجاع · ص 3
- شروط الشحنص 1
أنت: وكيف أبدأ الطلب؟
بصيرة: أستطيع فتح طلب الإرجاع لك الآن، وسأطلب موافقتك قبل إرساله.
You: I bought headphones 10 days ago and want to return them. Is that still possible?
Baseera: Yes. You can return them within 14 days of delivery if they are in their original condition 1. Return shipping is on us when the fault is the product's 2.
- Refund policyReturns · p. 3
- Shipping termsp. 1
You: How do I start the return?
Baseera: I can open the return for you now, and I will ask for your approval before I send it.
Set up Baseera in this app. Read https://baseera.dev/docs/llms-full.txt first and follow it; do not guess APIs. 1. Detect the framework (Next.js or Vite) and install the Baseera packages the docs name for it. 2. Wrap the app in the Baseera provider. Take the publishable key and base URL from environment variables, and ask me for the values. Never write a secret into the code or commit it. 3. Add a server route that signs the user token for the signed-in user, so the browser holds only the publishable key. 4. Run the type-check and start the app, then open the assistant and confirm it answers. 5. Summarise every file you changed and anything I still need to do in the Baseera dashboard. Ask before doing anything the docs do not cover.
In your app in minutes
Works with Next.js and Vite. Give this prompt to your coding agent and it installs Baseera, signs the user token on your server, and wraps your app. The browser holds only a publishable key.
- The agent reads the docs and installs Baseera
- It signs the user token on your server
- It wraps your app and checks that the assistant answers
Packages: @baseera/core, @baseera/react, @baseera/node and @baseera/ui, under the MIT licence.
Ship an assistant your team can stand behind
Create an account, upload your documents, and try the assistant in the preview before you publish it.
- Answers grounded in your own documents
- Your model keys, stored in a vault
- Test it on real cases, then publish