Knowi supports multiple AI model providers for powering features like InstantSights, AI Recommendations, AI Chat, Document AI, and more. By default, Knowi uses its built-in AI service. You can optionally configure external providers for enhanced capabilities.
Supported Providers
| Provider | Models | Key Features |
|---|---|---|
| Knowi AI (default) | Mistral 7B | Fully private, hosted within Knowi infrastructure. No data leaves the environment. |
| OpenAI | GPT-6.1 Terra, GPT-6 Astra, GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, GPT-5.5, GPT-4.1, GPT-4o, and until OpenAI shuts them down on December 11, 2026, GPT-5, GPT-5 Mini, GPT-5 Nano and o3 |
Large context windows, advanced reasoning capabilities. |
| Anthropic (Claude) | Claude Fable 5.2, Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5, Claude Fable 5.1, Claude Fable 5, Claude Opus 5, Claude Sonnet 5, Claude Opus 4.8, Claude Sonnet 4.6, Claude Opus 4.6, Claude Haiku 4.5 | Large context windows, strong analytical capabilities, streaming support, and configurable reasoning depth. |
| Google (Gemini) | Google Gemini 4 Carbon, Google Gemini 4 Argon, Google Gemini 3.7 Flash, Google Gemini 3.5 Flash Lite, Google Gemini 3.1 Pro, Google Gemini 2.5 Pro, Google Gemini 2.5 Flash, Google Gemini 2.5 Flash Lite | Large context windows, streaming support. |
| Azure AI | Azure OpenAI GPT-5.5, GPT-5.4 Mini, GPT-5, GPT-4o, GPT-4.1 | OpenAI models served from your own Azure OpenAI resource. Requires an endpoint URL in addition to an API key. |
| Amazon Bedrock | Claude Sonnet 5.5, Claude Opus 5.5, Claude Haiku 5.5, OpenAI GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Astra, GPT-6 Luna, xAI Grok 4.7, Amazon Nova 2 Lite | Models served from your own AWS account through the Bedrock Converse API. Requires a long-term Bedrock API key and its AWS region. |
Configuring AI Providers
To configure an external AI provider:
- Navigate to User Settings from the bottom-left sidebar.
- Open the AI Settings tab.
- Under the provider section, click CHANGE next to the provider you want to configure (OpenAI, Anthropic, Google, or Azure AI).
- Enter your API key in the modal and click Save. For Azure AI, enter your endpoint URL as well
- Once a key is saved, the corresponding models will appear in the Default AI Model Selection dropdowns for each AI feature.
Per-Feature Model Selection
Knowi allows you to select different AI models for different features. Under Default AI Model Selection in AI Settings, you can assign a provider for each of the following:
- NLP Search - Powers natural language queries across datasets.
- AI Chat - Powers the long-form AI chat experience.
- Insights - Powers InstantSights summaries.
- Recommendation Widget - Powers AI-generated recommendations.
- Documents - Powers Document AI question answering.
- Dashboards - Powers AI Dashboard generation.
- Agents - Powers Agentic AI assistants.
For example, you could use Knowi AI for NLP Search (to keep data private) while using Claude Sonnet 5.5 for AI Recommendations (for more advanced analysis).
Anthropic (Claude) Setup
- Obtain an API key from Anthropic Console.
- In Knowi AI Settings, click CHANGE next to the Anthropic (Claude) row.
- Enter your API key and save.
- Select a Claude model from the feature dropdowns.
Available Claude Models:
| Model | Best for | Reasoning Effort | Task Budget |
|---|---|---|---|
| Claude Fable 5.2 | The most demanding analysis and long-running agent work. Always reasons before answering. | Up to Max | Required |
| Claude Fable 5.1 | Superseded by Claude Fable 5.2. | Up to Max | Required |
| Claude Fable 5 | Superseded by Claude Fable 5.2. | Up to Max | Required |
| Claude Opus 5.5 | Complex analytical tasks and agent work. The strongest general-purpose choice. Always reasons before answering. | Up to Max | Optional |
| Claude Opus 5 | Superseded by Claude Opus 5.5. | Up to Max | Optional |
| Claude Sonnet 5.5 | Balanced performance and speed. The default Claude model when none is selected. Always reasons before answering. | Up to Max | Optional |
| Claude Sonnet 5 | Superseded by Claude Sonnet 5.5. | Up to Max | Not supported |
| Claude Opus 4.8 | Superseded by Claude Opus 5.5. | Up to Max | Optional |
| Claude Sonnet 4.6 | Superseded by Claude Sonnet 5.5. | Up to High | Not supported |
| Claude Opus 4.6 | Superseded by Claude Opus 5.5. | Up to High | Not supported |
| Claude Haiku 5.5 | Fastest and lowest-cost model. Best for high-volume, lower-complexity tasks. 1M context. | Up to Max | Optional |
| Claude Haiku 4.5 | Superseded by Claude Haiku 5.5. | Not supported | Not supported |
When a model is retired. Once the vendor (Anthropic, OpenAI or AWS) retires a model, Knowi stops listing it, and every feature still set to it switches automatically to the newest model that replaces it (for example, a retired Claude Fable model is replaced by Claude Fable 5.2). Your saved settings are not rewritten: the replacement shows as selected the next time you open AI Settings, and it keeps the reasoning effort and task budget saved for the retired model until you next save that feature's models. Models the vendor has scheduled for retirement are marked Deprecated with their retirement date.
Reasoning Effort
Claude models can spend more or less time reasoning before answering. Set Reasoning Effort in the Anthropic (Claude) section of AI Settings to control that tradeoff across every feature using a Claude model:
- Low - Fastest and cheapest. Best for short, well-scoped tasks.
- Medium - Reduces token usage where the task does not need deep reasoning.
- High - The default when no effort is set (Claude Opus 5.5 defaults to Medium). A good balance for most analytical work.
- Extra High - Recommended for the hardest analysis and agent tasks.
- Max - Highest capability, highest cost. Best when correctness matters more than speed.
Leave the setting empty to use each model's default. If you select an effort level a model does not support, Knowi automatically steps down to the closest supported level rather than failing the request ? Extra High becomes High on Claude Sonnet 4.6 and Claude Opus 4.6, and effort is skipped entirely on Claude Haiku 4.5.
Task Budget
Task Budget caps how many tokens a single Claude request may spend, including reasoning. Set it in tokens in the Anthropic (Claude) section of AI Settings. The minimum accepted value is 20,000 tokens; values below that, or non-numeric entries, are ignored and the request runs without a budget.
Because the Claude Fable models are premium-priced and always reason before answering, Knowi requires a task budget before Claude Fable 5.2, Claude Fable 5.1 or Claude Fable 5 can be used. None of them appears in the feature dropdowns until a budget is set. If the budget is later cleared, a Fable model already saved for a feature fails with a message asking you to set a budget or choose a different model. Every other model, including Claude Opus 5.5 and Claude Sonnet 5.5, treats the budget as optional.
Google (Gemini) Setup
- Obtain an API key from Google AI Studio.
- In Knowi AI Settings, click CHANGE next to the Google (Gemini) row.
- Enter your API key and save.
- Select a Gemini model from the feature dropdowns.
Available Gemini Models:
| Model | Best for |
|---|---|
| Gemini 4 Carbon | The deepest reasoning in the Gemini family, spending more thinking on each turn. A strong fit for complex analytical work and questions that depend on industry knowledge. Also well suited to code and security-oriented tasks. |
| Gemini 4 Argon | Faster responses and lower cost per turn than Gemini 4 Carbon, at the same per-token price. A good default for most analytical work on the Gemini 4 generation. |
| Gemini 3.7 Flash | Balanced speed and capability. The default Gemini model when none is selected. |
| Gemini 3.5 Flash Lite | Fastest and lightest. Best for high-volume, lower-complexity tasks. |
| Gemini 3.1 Pro | Superseded by Gemini 4 Argon. |
| Gemini 2.5 Pro | Superseded by Gemini 4 Argon. |
| Gemini 2.5 Flash | Superseded by Gemini 3.7 Flash. |
| Gemini 2.5 Flash Lite | Superseded by Gemini 3.5 Flash Lite. |
The Gemini 2.5 models and Gemini 3.1 Pro remain selectable so existing configurations keep working, but new setups should use the Gemini 4 models, Gemini 3.7 Flash or Gemini 3.5 Flash Lite. Google has already restricted some 2.5 models to accounts that used them previously; if a 2.5 model returns an error saying it is no longer available to new users, switch to the model listed above as its replacement.
Gemini 4 Carbon and Gemini 4 Argon always reason before answering, and there is no setting to turn that off. The reasoning counts toward the response length limit, so Knowi allows these two models up to 16,000 output tokens per response instead of the shorter limit used for other Gemini models. Gemini 3.7 Flash remains the default Gemini model when none is selected.
Gemini models can be used for every AI feature, including Document AI. On-premises deployments need a Knowi AI service build that includes Gemini support before Document AI can answer over a Gemini model; contact your Knowi administrator if a Gemini selection works elsewhere but not in Document AI.
Note: Gemini models are subject to Google's own per-model availability and quota. On the Google free tier some models return no output at all rather than an error. When a request to an external provider fails, Knowi now surfaces the provider's own error message rather than a generic one, so the reason - invalid key, model unavailable, quota exhausted - is visible directly in the feature you were using. Internal Knowi AI errors continue to show the generic message, since they are not actionable by the user.
OpenAI Setup
- Obtain an API key from OpenAI Platform.
- In Knowi AI Settings, click CHANGE next to the OpenAI row.
- Enter your API key and save.
- Select an OpenAI model from the feature dropdowns.
Available OpenAI Models:
| Model | Best for | Reasoning Effort |
|---|---|---|
| GPT-6.1 Terra | Balancing capability against cost. Good default for most use cases. | Low to High |
| GPT-6 Astra | OpenAI's most capable model, built for the hardest end-to-end work. | Low to High |
| GPT-5.6 Sol | Complex reasoning and the most demanding analysis. | Low to High |
| GPT-5.6 Terra | Superseded by GPT-6.1 Terra. | Low to High |
| GPT-5.6 Luna | Cost-sensitive, high-volume work where speed matters most. | Low to High |
| GPT-5.5 | Strong general-purpose reasoning, previous generation. | Low to High |
| GPT-5 | Deprecated: OpenAI shuts it down on December 11, 2026. Replaced by GPT-5.6 Sol. | Low to High |
| GPT-5 Mini | Deprecated: OpenAI shuts it down on December 11, 2026. Replaced by GPT-6.1 Terra. | Low to High |
| GPT-5 Nano | Deprecated: OpenAI shuts it down on December 11, 2026. Replaced by GPT-5.6 Luna. | Low to High |
| GPT-4.1 | General-purpose work with a large context window. | Not supported |
| GPT-4o | General-purpose work. | Not supported |
| o3 | Deprecated: OpenAI shuts it down on December 11, 2026. Replaced by GPT-5.6 Sol. | Low to High |
OpenAI Effort
OpenAI reasoning models can spend more or less time thinking before answering. Set OpenAI Effort in the OpenAI section of AI Settings to control that tradeoff across every feature using an OpenAI model:
- Low - Fastest and cheapest. Best for straightforward, high-volume work.
- Medium - Balanced reasoning depth against speed and cost.
- High - Deepest reasoning. Best for complex analysis, at the cost of more tokens and more time.
Deprecated OpenAI Models
OpenAI shuts down GPT-5, GPT-5 Mini, GPT-5 Nano and o3 on December 11, 2026. They are marked Deprecated in AI Settings and remain selectable until then, but new setups should use their replacements: GPT-5.6 Sol for GPT-5 and o3, GPT-6.1 Terra for GPT-5 Mini, and GPT-5.6 Luna for GPT-5 Nano. o3-mini was shut down on October 23, 2026 and is no longer listed. For the current schedule, see OpenAI's deprecations page.
On the shutdown date, Knowi removes the model from the model lists, and every feature still set to it switches automatically to its replacement, so nothing falls back to a different provider. Your saved settings are not rewritten: the replacement shows as selected the next time you open AI Settings.
Azure AI Setup
- In the Azure portal, open your Azure OpenAI resource and go to Keys and Endpoint to copy your API key and endpoint URL (for example,
https://myresource.openai.azure.com). - Deploy the models you want to use in your Azure resource, using their standard names:
gpt-5.5,gpt-5.4-mini,gpt-5,gpt-4o,gpt-4.1. - In Knowi AI Settings, click CHANGE next to the Azure AI row.
- Enter your API key and endpoint URL, then save.
- Select an Azure model from the feature dropdowns.
Note: Unlike other providers, Azure AI requires both an API key and an endpoint URL. Models must be deployed in your Azure resource before Knowi can use them; if a model is not deployed, requests will fail with a deployment not found error.
Amazon Bedrock Setup
Amazon Bedrock runs the models in your own AWS account. Knowi calls them through the Bedrock Converse API with a Bedrock API key, so prompts and data go to your AWS account rather than to the model vendor.
- In the Amazon Bedrock console, go to API keys and create a long-term API key. Short-term keys are not supported: they expire within 12 hours and Knowi has no way to refresh them.
- Enable each model you plan to use under Model access in the Bedrock console. Third-party models are enabled through an AWS Marketplace subscription. Until a model is enabled, requests to it fail with an access denied error.
- Make sure the key's IAM identity allows
bedrock:InvokeModelandbedrock:InvokeModelWithResponseStreamon the inference profile and on the underlying foundation-model ARNs. GPT-6 models and Grok 4.7 also needbedrock:InvokeModelon your account's default project:arn:aws:bedrock:<region>:<account-id>:project/default. - In Knowi AI Settings, click CHANGE next to the Amazon Bedrock row. Enter the API key, the AWS region to use (where you enabled the models, for example
us-east-1), and the data residency, then save. The Endpoint field is optional: fill it in only to reach Bedrock through a VPC (PrivateLink) endpoint. - Select a Bedrock model from the feature dropdowns. Only models that Bedrock offers from your region with the chosen data residency are listed.
Data Residency
Bedrock serves these models through cross-Region inference profiles. Knowi picks the profile from your region and the data residency setting:
-
Geographic (default): requests stay within your region's geography, for example the US profile from
us-east-1orca-central-1, or the EU profile fromeu-central-1. - Global: requests can be routed to any AWS region that serves the model. Use it when you have no residency requirement, or when a model has no geographic profile for your region.
Bedrock Models
| Model | Bedrock model ID | Geographic profiles | Notes |
|---|---|---|---|
| Claude Sonnet 5.5 | anthropic.claude-sonnet-5-5 |
US, EU, AU | Recommended default. 1M context. |
| Claude Opus 5.5 | anthropic.claude-opus-5-5 |
US, EU, AU, JP | Hardest tasks. 1M context. |
| Claude Haiku 5.5 | anthropic.claude-haiku-5-5 |
US, EU, AU, JP | Fast and low cost. 1M context. |
| Claude Fable 5.2 | anthropic.claude-fable-5-2 |
US | Most capable. Requires the aws_review data retention mode (see below). 1M context. |
| OpenAI GPT-6.1 Sol | openai.gpt-6.1-sol |
US | 1M context. |
| OpenAI GPT-6.1 Terra | openai.gpt-6.1-terra |
US | Mid-price, between Luna and Sol. Long context (1.05M). |
| OpenAI GPT-6 Astra | openai.gpt-6-astra |
US | Long context (1.05M). |
| OpenAI GPT-6 Luna | openai.gpt-6-luna |
US | Lower cost. |
| xAI Grok 4.7 | xai.grok-4.7 |
US | Reasoning model; 524K context shared between input and output. |
| Amazon Nova 2 Lite | amazon.nova-2-lite-v1:0 |
US, EU, JP | Amazon's own low-cost model. |
Every model is also available with Global data residency.
Reasoning Effort on Bedrock
The Claude models on Bedrock use the same Effort setting as Anthropic (Claude) models; task budgets are not available on Bedrock. The GPT-6 models and Grok 4.7 use the OpenAI reasoning effort setting (Low, Medium or High; Medium unless you choose otherwise). Nova 2 Lite runs without reasoning and has no effort setting.
Quotas
New AWS accounts have low default tokens-per-minute quotas for Bedrock, and GPT-6 models count each output token as 10 tokens of quota. Under load, expect "rate limit exceeded" errors until you raise the quotas in the AWS Service Quotas console.
Claude Fable 5.2 on Bedrock
Claude Fable 5.2 answers only once your AWS account's Bedrock data retention mode is aws_review (or the legacy provider_data_share) in the region you use. In that mode AWS retains prompts and responses for up to 30 days for abuse detection and may review flagged ones; the content is not shared with Anthropic. Until it is set, every request fails with a message saying so. Set it once per region with the Bedrock control plane API, using the same Bedrock API key:
curl -X PUT https://bedrock.us-east-1.amazonaws.com/data-retention \
-H "Authorization: Bearer $AWS_BEARER_TOKEN_BEDROCK" \
-H "Content-Type: application/json" \
-d '{ "mode": "aws_review" }'Only the US geographic profile exists for Fable 5.2; from regions outside the US, choose Global data residency. Task budgets are not available on Bedrock, so Fable 5.2 on Bedrock does not require one. It also declines more requests than other Claude models.
Bedrock Models Not Offered
- Claude Fable 5.1 and Claude Fable 5: superseded by Claude Fable 5.2.
- Kimi K3: its Bedrock model card advises against the Converse API, which Knowi uses for every Bedrock model, and it rejects PDF and HTML documents on Converse.
Bedrock on Premise
- The Knowi application server needs outbound HTTPS to
bedrock-runtime.<region>.amazonaws.com. For Document AI, the AI service container needs the same access, because Document AI answers are generated there. Document embeddings stay on Knowi's own model and are never sent to Bedrock. - To use a VPC (PrivateLink) endpoint or another Bedrock host, enter its base URL in the optional Endpoint field of the Amazon Bedrock key, for example
https://vpce-0123.bedrock-runtime.us-east-1.vpce.amazonaws.com. Leave it blank for the public regional endpoint. The same URL is used by the AI service container, so it must resolve from there as well. - If outbound traffic goes through a proxy, set the JVM proxy properties (
-Dhttps.proxyHost,-Dhttps.proxyPort,-Dhttp.nonProxyHosts) for the application server, andHTTPS_PROXY/NO_PROXYfor the AI service container.