Configuring the Chat Agent Innovation Release

The EDB Hybrid Manager (HM) Chat Agent is delivered with HM and powered by one or more large language models (LLMs) that you provide. As an HM administrator, you supply the connection details for models that already run elsewhere. After that, anyone in your organization can use Chat Agent from the HM console.

This page is for HM administrators. To learn how to use Chat Agent, see the overview.

Before you begin

Confirm these prerequisites:

  • HM 2026.11 or later is installed and reachable.
  • The chatbot scenario is enabled in your HM deployment. The sparkle icon stays hidden when the scenario isn't installed. To enable it, list chatbot under scenarios in your HybridControlPlane manifest. See HM installation scenarios for details.
  • You have the connection details for at least one LLM, either a third-party API or a self-hosted endpoint. See Supported model providers.
  • The LLM supports a context window of at least 128K tokens. Chat Agent loads skill instructions, tool schemas, and conversation history on every turn. Models with smaller context windows don't have enough context.
  • You can run edbctl against the Kubernetes cluster that hosts HM. The edbctl command needs --version set to 2026.11 or later.
Chat Agent in earlier releases

In HM 2026.9 and earlier, Chat Agent was part of the ai scenario. It used HM AI Factory model serving and inference services for its models. HM 2026.10 removed the ai scenario, the model serving that came with it, and Chat Agent. From 2026.11, Chat Agent returns as its own chatbot scenario. It doesn't need the ai scenario, HM AI Factory, or an inference service. It connects directly to the endpoints you configure.

Important

Chat Agent answers questions about your live HM estate. Choose a provider whose data-handling policy matches your organization's requirements. Self-hosted models keep prompts and responses inside your infrastructure.

Configuring the models

Configure Chat Agent models as part of the HM installation, with edbctl. For each model, you provide a unique ID, the model name, and the endpoint URL. An API key is optional, so endpoints that don't need a key work too.

  1. Gather these details for each model. The examples use Google Gemini:

    DetailDescriptionExample
    IDA unique identifier that you choose for the model.gemini-flash
    NameThe model name that the endpoint expects.gemini-3.5-flash-lite
    URLThe base URL of the endpoint.https://generativelanguage.googleapis.com
    API keyThe key for the endpoint. Optional for endpoints that don't need one.Your Gemini API key
    API typeopenai (the default), anthropic, or gemini. See Supported model providers.gemini
  2. Create the model secret. The <version> must be 2026.11 or later. Run the command and answer the prompts.

    edbctl hybrid-manager create-install-secrets --version <version> --scenario chatbot

    For automated installations, pass the same details as EDB_CHATBOT_MODEL_* environment variables. See Customizing Chat Agent model secrets.

  3. Install HM with chatbot in spec.scenarios.

Adding more than one model

Chat Agent supports several models at the same time. Chat Agent lists every configured model in the model picker, and the first model is the default. For example, configure a stronger model for cluster management and a lighter model for documentation questions.

To add more than one model, do one of the following:

  • In interactive mode, answer yes when edbctl asks whether to add another model, and enter the details for each one.
  • With environment variables, number the settings by position. Use EDB_CHATBOT_MODEL_ID1 and EDB_CHATBOT_MODEL_ID2, and so on. Each model needs a unique ID.

For a complete example with two Gemini models, see Adding more than one model.

Changing the models after installation

To add, replace, or remove models, run the same edbctl command again with the --force flag and the complete list of models you want. The new list replaces the old one, including capability settings. HM restarts Chat Agent automatically to pick up the change. See Changing the models later.

Setting model capabilities

Chat Agent detects the strength of recognized models automatically. You can override detection, or describe a model that isn't recognized, with these optional settings. Provide them when you create the model secret, as prompts or EDB_CHATBOT_MODEL_<SETTING><N> variables, where <N> is the model's position.

SettingVariable suffixValuesPurpose
TierTIERS, M, L, or XLOverrides the detected strength tier.
Context windowCONTEXT_WINDOWA number followed by k or m, for example 128kTells Chat Agent how much context the model supports.
Thinking levelsTHINKING_LEVELSAny of low, medium, high, xhigh, and max, separated by commasLists the reasoning levels the model accepts.
Default thinking levelDEFAULT_THINKING_LEVELOne of the thinking levels you listedSets the model's default reasoning effort.

Set thinking levels only on models that support thinking or reasoning modes. When you set thinking levels, you must also set a default, and the default must be one of the listed levels. If you don't set them, the model uses its own default thinking level. That default works well for most tasks. Raise it for more complex tasks, or lower it to speed up responses.

Warning

Some reasoning models don't accept a per-request thinking-level setting. On those models, setting a thinking level causes the model to return an error and Chat Agent stops responding. Confirm the model accepts your chosen level before you set it.

Set a tier only when your model isn't on the recognized list. See Assigning a tier to an unrecognized model for the procedure.

Verifying your configuration

  1. In the HM console, sign in as a user that belongs to at least one project.
  2. Select the sparkle icon at the top right.
  3. Confirm the panel opens with the header HM AI and the placeholder Ask me anything... in the input.
  4. If you configured more than one model, select the model picker and confirm your models appear.
  5. Send a quick prompt, such as: What can you do?
  6. Confirm a streamed response begins within a few seconds.

If the sparkle icon is missing or the panel can't be opened, see Troubleshooting.

Accessing the Chat Agent

Chat Agent is available on every authenticated page in the HM console.

  1. Look for the rainbow sparkle icon at the top right of the HM header.
  2. Select the icon to open Chat Agent.

If the icon is missing or disabled, see Troubleshooting.

Chat Agent opens in two layouts:

  • Full screen — The default. The conversations list appears on the left.
  • Floating panel — A compact, draggable panel docked at the bottom right of the HM console. Use the expand or shrink button to switch between modes.

Using the Chat Agent

Ask a question

  1. Select Model from the picker if your administrator configured more than one. The model determines which capabilities are available — see Model-dependent capabilities.
  2. Type your question in the Ask me anything input and press Enter.
  3. Use Shift+Enter to add a line break within your message.
  4. Chat Agent streams a response in the chat window. Documentation answers include inline citations so you can open the source page directly.
  5. Select the stop button to interrupt generation early.

If Chat Agent needs more information before completing your request — for example, which project to target or a connection password — it pauses and presents a structured form. Fill in the form to continue, or cancel it (or type an unrelated message) to dismiss it.

Use page context

When you open Chat Agent from a cluster, project, or other resource page, it picks up that page's context automatically. The project and cluster IDs appear as small tags above the input, so you can ask about that resource without naming it — for example, Show me the active alerts applies to the cluster you're currently viewing.

Manage conversations

Each conversation is saved to your HM organization so you can return to it later. Select New Chat to start a fresh conversation, or open the sidebar with the menu icon to return to an earlier one.

Chat Agent gives each new conversation a short auto-generated title. From the sidebar you can also Rename a conversation or Delete one — deletion can't be undone.

For conversation privacy details, see Conversation storage and isolation.

Recommendations

  • Start with a Large model — Large is the lowest tier that covers Postgres Cluster lifecycle management with structured forms and the full set of Medium-tier capabilities. gemini-2.5-pro, gpt-5, and gpt-oss-120b sit in this tier.
  • Pick a self-hosted model if your prompts contain regulated or proprietary data — This choice keeps Chat Agent prompts and responses inside your infrastructure.
  • Configure more than one model to let users pick a tier per conversation — For example, add a Small model for documentation lookups and a Large model for cluster-creation flows.

Supported model providers

Chat Agent communicates with models through three API protocols. Any provider — third-party or self-hosted — that speaks one of these protocols works. Set the protocol with the API type when you create the model secret. For Gemini, use the base URL https://generativelanguage.googleapis.com. For OpenAI-compatible endpoints, include the version path, for example https://llm.example/v1. EDB doesn't build or ship AI models.

API type and protocolCompatible providers and runtimes
openai (OpenAI v1)OpenAI, Azure OpenAI, NVIDIA NIM containers, Amazon Bedrock (via OpenAI-compatible gateway), vLLM, and other OpenAI-compatible self-hosted servers
anthropic (Anthropic Messages)Anthropic API, Anthropic-compatible gateways and proxies that preserve the Messages API surface
gemini (Google Gemini)Google AI Studio (Gemini API), Vertex AI with the Gemini API surface

Model strength tiers

Chat Agent classifies every configured model into a strength tier — Small, Medium, Large, or Extra Large — based on the model's capability. A model's tier determines which Chat Agent skills are available when that model is selected.

Strength tiers

The tiers correspond to approximate model scale and capability:

  • Small — Roughly 7–30B parameter class.
  • Medium — Roughly 30–100B parameter class.
  • Large — gpt-4.1, gemini-2.5-pro, gpt-5, and gpt-oss-120b class.
  • Extra Large — Current flagship models: the Claude 4.5+ family (including claude-haiku-4-5), gpt-5.6 and later, gpt-5.3-codex, gemini-3.1-pro-preview, and comparable frontier models.
Note

claude-fable-5 requires the underlying inference provider account to have 30-day (nonzero) data retention enabled. This is a platform- or account-level setting (Amazon Bedrock, Google Vertex AI, Microsoft Foundry, or the direct Anthropic API) outside of HM's control. Accounts configured for zero data retention will see every claude-fable-5 request fail with a 400 error regardless of HM configuration.

A model that isn't on the recognized list below defaults to Small.

Recognized models

ModelTier
gpt-4o-miniSmall
gpt-oss-20bSmall
gpt-5.4-nanoMedium
gpt-4.1-miniMedium
gpt-5.4-miniLarge
gpt-4.1Large
gpt-4oSmall
gpt-oss-120bLarge
gpt-5.6-lunaExtra Large
gpt-5.6-solExtra Large
gpt-5.6-terraExtra Large
gpt-5.5-proExtra Large
gpt-5.5Extra Large
gpt-5.4-proExtra Large
gpt-5.4Extra Large
gpt-5.2Extra Large
gpt-5.3-codex-sparkExtra Large
gpt-realtime-2.1Extra Large
o1Extra Large
o1-proExtra Large
o3-miniLarge
gpt-6-astraExtra Large
gpt-6-lunaExtra Large
gpt-6.1-solExtra Large
claude-haiku-4-5Extra Large
claude-sonnet-4-5Extra Large
claude-sonnet-4-6Extra Large
claude-opus-4-5Extra Large
claude-opus-4-6Extra Large
claude-opus-4-7Extra Large
claude-opus-4-8Extra Large
claude-sonnet-5Extra Large
claude-fable-5Extra Large
claude-opus-5Extra Large
claude-fable-5-1Extra Large
claude-haiku-5-5Extra Large
claude-opus-5-5Extra Large
claude-sonnet-5-5Extra Large
gemini-3.1-flash-liteSmall
gemini-3-flash-previewLarge
gemini-3.1-pro-previewExtra Large
gemini-3.5-flashExtra Large
gemini-3.5-flash-liteLarge
gemini-3.6-flashExtra Large
gemini-3.7-flashExtra Large
deep-research-max-preview-04-2026Extra Large
deep-research-preview-04-2026Extra Large
gemini-2.5-computer-use-preview-10-2025Large
gemini-3.1-flash-lite-imageExtra Large
gemini-3.1-flash-live-previewLarge
gemini-3.1-pro-preview-customtoolsExtra Large
gemma-4-26b-a4b-itSmall
gemma-4-31b-itSmall
gemini-3.8-flashLarge
gemini-4-argonExtra Large
nvidia/llama-3.3-nemotron-super-49b-v1Small
nvidia/llama-3.3-nemotron-super-49b-v1.5Small
nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16Small
nvidia/Nemotron-Cascade-2-30B-A3BSmall
nvidia/llama-3.1-nemotron-ultra-253b-v1Large
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16Large
nvidia/cosmos-reason2-8bSmall
nvidia/llama-3.1-nemotron-70b-instructMedium
nvidia/llama-3.1-nemotron-nano-8b-v1Small
nvidia/llama-3.1-nemotron-nano-vl-8b-v1Small
nvidia/nemotron-3-ultra-550b-a55bMedium
nvidia/nemotron-3.5-lightning-30b-a3bSmall
nvidia/nemotron-mini-4b-instructSmall
nvidia/nemotron-nano-12b-v2-vlSmall
nvidia/nemotron-voicechatSmall
nvidia/nvidia-nemotron-nano-9b-v2Small
nvidia/nemotron-3-nano-30b-a3bSmall
nvidia/nemotron-3.5-super-vl-120b-a12bSmall
deepseek-ai/deepseek-v4-flashExtra Large
deepseek-ai/deepseek-v4-proExtra Large
deepseek-ai/deepseek-v4.1-flashExtra Large
qwen/qwen3-next-80b-a3b-instructMedium
qwen/qwen3-next-80b-a3b-thinkingMedium
qwen/qwen3.5-122b-a10bLarge
qwen/qwen3.5-397b-a17bExtra Large
qwen/qwen3-coder-480b-a35b-instructExtra Large
meta/llama-4-maverick-17b-128e-instructExtra Large
meta/llama-3.2-11b-vision-instructSmall
meta/llama-3.2-1b-instructSmall
meta/llama-3.2-90b-vision-instructMedium
meta/muse-glimmer-30bMedium
meta/llama-3.1-70b-instructMedium
minimaxai/minimax-m2.7Extra Large
minimaxai/minimax-m3Extra Large
z-ai/glm5Extra Large
z-ai/glm-5.1Extra Large
z-ai/glm-5.2Large
z-ai/glm-5.3-flashExtra Large
z-ai/glm-5.3Extra Large

The following aliases are recognized as the same canonical model:

  • Anthropic models on Amazon Bedrock — Cross-region prefixes (us., eu., jp., au., global.) and provider-prefixed variants (anthropic.<model>-<date>-v1:0).
  • OpenAI gpt-oss-20b and gpt-oss-120b — Provider-prefixed variants openai/<model>, openai.<model>-1:0, gpt-oss:<size>.

Assigning a tier to an unrecognized model

Models whose names don't match any row in the table above (and aren't covered by the alias rules) default to Small. Many of the more advanced skills are unavailable as a result.

To override this setting, set the model's tier when you create the model secret. Use the TIER prompt or the EDB_CHATBOT_MODEL_TIER<N> variable with one of these values:

Tier valueTier
SSmall
MMedium
LLarge
XLExtra Large

Use the value that best matches your model's actual capability. After you update the secret and Chat Agent restarts, it honors the tier you set instead of defaulting to Small.

Skills available at each tier

A skill is available when the selected model's tier is at or above the skill's minimum tier. Higher tiers add capabilities without removing the existing ones.

Skill / capabilityMinimum tier
HM documentation Q&A (concepts, architecture, how-to, with citations)Small
Browse and manage estate, projects, organization users, project users, organization roles, and project rolesMedium
Project and organization activity feedsMedium
Iceberg and Lakehouse catalog inventory (list and inspect, read-only)Medium
Kubernetes workload observability — Grafana data sources, Prometheus queries, Loki queriesMedium
Cluster realtime and historical metrics, charts, and available metric metadataMedium
Active alerts (per cluster, per project), aggregated health scoresMedium
Configuration, index, statistics, and security recommendationsMedium
Data Migration Service status — sources, destinations, mappings, errors, assessmentsMedium
Oracle-to-Postgres SQL translationMedium
Resource-usage analytics (compute, memory, and storage consumption)Medium
Postgres Cluster lifecycle managementMedium
AI Factory inventory — models, inference services, knowledge bases, pipelines, and flowsLarge
Image library and deployable Postgres versions per locationLarge
Read-only database, schema, and cluster-configuration inspectionLarge
Cluster health reports — list, fetch, analyze, compareLarge
Postgres Cluster lifecycle management with structured follow-up questionsLarge
Schema design recommendations driven by top running queriesExtra Large

A Small model offers only HM documentation Q&A. Postgres Cluster lifecycle management starts at the Medium tier, with structured follow-up questions added at Large and above. Cluster health reports need a Large model, and schema design recommendations need an Extra Large model. For a full breakdown of what each tier supports, see What the Chat Agent can do.

Disabling Chat Agent for an organization

If the chatbot scenario is installed but no model is configured, the sparkle icon is shown but disabled. A tooltip tells users that the administrator needs to configure an available model.

To remove the sparkle icon entirely, leave chatbot out of spec.scenarios at install time. See HM installation scenarios.

Next steps