Your Intelligent Space.

Architecture at a glance

The browser application coordinates several independent services. Each integration has its own key, endpoint, quota and availability.

Model card

A concise description of the system boundary, intended use and limitations.

Purpose & intended users

Zorae is designed as a general-purpose assistant workspace for people who want to chat with a configured AI provider, work with supported files, optionally retrieve web information, explore places and manage local conversation context.

System boundary

Zorae does not independently train or own the underlying language models. Model behavior, knowledge cutoffs, safety filters, context limits, pricing and uptime are determined by the selected upstream provider and model.

Known limitations

• Generated responses can be incorrect, incomplete, outdated or fabricated. Verify consequential claims with primary sources.
• Tool results depend on configured API keys, quotas, network access, CORS/browser policies and third-party service availability.
• File formats, payload size, image understanding and audio capabilities vary by provider and browser.
• Local browser storage is not cloud synchronization or a guaranteed backup. Clearing site data can remove records.
• The assistant may reflect limitations, biases and safety policies of the selected upstream model.

First setup

1. Open Settings from the left sidebar and choose APIs & Models.
2. Select the provider you want to use. Each provider has a separate API-key field; keys are not interchangeable.
3. Choose a model available in the selector, enter the corresponding key and save.
4. Return to chat and send a short test prompt. The active model is shown in the lower-left profile area.
5. If you want web search, add a Tavily key in its dedicated field. Geographic lookups use OpenStreetMap/Overpass integration.

Chat modes, memory & response depth

Response-depth modes

Fast, Balanced and Deep are application-level response modes. They adjust the response behavior/instructions; they are not separate model providers and do not guarantee a fixed latency or reasoning quality.

Files, messages & local data

• Use the attachment control to add supported files. Text content can be extracted into the prompt context; images are sent only where the selected provider path supports them.
• Provider-specific file limits and supported formats apply. Large files may exceed browser memory or API request limits.
• Message actions include copy, edit and regenerate. Editing a user message replaces that message and subsequent conversation turns.
• Code blocks include a dedicated copy action. Clipboard access may require HTTPS or localhost and browser permission.
• The app includes JSON backup export/import for chat and memory data. Review the file before sharing: it may contain personal conversation content.
• Storage diagnostics and reset controls are available in settings. Clearing browser/site data may permanently remove local records.

Web search & place discovery

Tavily web search

1. Open Settings → APIs & Models and save a valid Tavily API key.
2. Zorae may search the web automatically when it determines that current information, recent sources or verification would help.
3. When results are returned, titles, URLs and summaries are added to model context and source links can appear with the answer.

A response without retrieved source links does not prove that a web search was performed. Check the key, account quota, network and browser console if results are missing.

OpenStreetMap place lookup

Place discovery uses OpenStreetMap data through the Overpass API. Coverage and tagging vary by region; map records may be incomplete or out of date. Treat results as discovery leads and confirm critical details with the venue or official source.

Audio capabilities

The source contains provider-specific audio/transcription pathways. The available route depends on provider configuration, key, request format and upstream model availability. Audio identifiers are not part of the 12 selectable chat-model IDs listed below; do not substitute them into the chat model selector.

Prompt engineering

A useful prompt states the task, context, constraints and desired output. Iterate with evidence rather than assuming the first answer is correct.

Reusable prompt structure

Task: [what you need]
Context: [relevant background and source material]
Constraints: [scope, audience, length, exclusions]
Output: [format, structure, language]
Verification: [what must be cited or checked; state uncertainty]

Illustrative examples

Models & providers

The following identifiers are the configured identifiers confirmed for this Zorae build. They are reproduced as supplied; this documentation does not independently assert that every identifier is a public catalog model name at the provider.

Claude fallback: when the selected model is claude-opus-5-5, the source retries with claude-fable-5-2 only when the requested model appears unavailable (including HTTP 404/model-not-found indications). It does not use this fallback for authentication, rate-limit, network or general server errors. The fallback is not a selectable model.

Pricing & response time

No provider price or latency benchmark is asserted for these configured identifiers. Exact cost depends on the provider's actual model mapping, input/output tokens, modalities, region and current pricing schedule. Latency varies with provider load, network, prompt size, streaming and tool use. Check the official pricing pages below before estimating operating cost.

API errors & recovery

HTTP status meanings are common conventions. The provider's response body is authoritative for the specific cause; not every provider uses every status identically.

Data, privacy & responsible use

This is a technical summary, not legal advice and not a substitute for Zorae's future Terms of Use or Privacy Policy.

What the application stores locally

The source uses browser-local persistence for application records such as chats, private chats, memories and settings. Local storage is tied to the browser profile and device. Backup export can create a file containing conversation data.

What happens when you call a provider

Prompts and attached content sent to Gemini, OpenAI, Claude, Qwen, Tavily or other configured services are processed under those services' own terms, privacy controls and retention policies. Zorae cannot guarantee the provider's training or retention behavior. Review each provider's current privacy documentation and account settings.

Training data: what can and cannot be said

Zorae itself is an orchestration application, not a separately trained foundation model. The underlying providers develop their own models using data and processes described in their respective public materials. Those disclosures differ by provider, product and model; this guide does not claim a single shared dataset or infer the exact training corpus of any configured model.

LGPD & GDPR — short orientation

• Avoid submitting personal or sensitive data unless necessary and authorized.
• Use private/temporary modes as application-level history controls, not as a guarantee of zero processing or retention by external providers.
• Understand the lawful basis, purpose, minimization, security, retention and data-subject rights relevant to your deployment.
• For requests involving access, correction, deletion or portability, consult the service provider and the responsible Zorae operator.

FAQ & troubleshooting

The model does not respond. What should I check?

Confirm the active provider, model selection and matching API key. Check account quota/billing, browser network errors, provider status and whether the model ID is accepted by that provider account.

Why does web search return no sources?

Confirm the Tavily key is saved and valid, the account has quota, the prompt explicitly asks for research and network/CORS rules allow the request. A generated answer without returned source links is not a sourced search result.

Where are my chats stored?

In the current browser profile's local database (IndexedDB in the normal path). Clearing site data, using another browser profile or switching devices can make the records unavailable.

Why did my private or temporary conversation disappear?

Private and temporary modes have different persistence rules from normal chats. Temporary chats are designed not to appear in recent history; private chats are isolated from normal history.

Why is a file or image not processed?

Check format, size, provider support, API request limits and browser memory. Try a smaller file and confirm the selected provider's current multimodal capabilities.

How do I restore my backup?

Use the import backup control in settings and select a Zorae JSON backup. Import merges supported chats and memories with current local records; inspect the file first and do not import untrusted content.

How do I get the latest changelog?

Open Zorae, use the left sidebar and click “Updates”. This documentation intentionally does not reproduce a changelog.

Updates

To see the current product update notes, open Zorae, use the left sidebar and click “Updates”. No release history is reproduced on this documentation page.