Configure model services
Model configurations are shared resources that agents can reuse. Configure and test the service first, then select it in an agent. A model response and a successful DAM governance tool call are separate acceptance checks.
LLM (large language model) configurations support text understanding and generation. Text Embedding configurations convert text into numeric vectors for tasks such as semantic retrieval; they do not generate conversation answers. Token limits count model text units, not characters. Vector dimensions specify the number of numeric values returned and must match the actual service.
Before you start
Obtain the protocol, API base URL, server model identifier, and credentials from the service owner. Confirm network access from DIG and whether the task needs a generative or embedding model. A web chat URL is not an API base URL.
Add a configuration
- Open Configuration Management > Model Configuration > Add Model.
- Enter a configuration name and choose model type and provider. The provider represents the protocol, not just the model brand.
- Enter the base URL and an API key if required.
- Allow automatic discovery or click Discover Models. Select the target model, or enter its exact server identifier if discovery is unavailable.
- Set type-specific parameters and optionally make it the system default.
- Click Test Connection. The current form requires a successful test before Save.
- Verify the saved name, provider, and model identifier, then test it in the intended agent.
Use Cancel or Close to discard changes. Discovery does not replace a connection test or task validation.
Recommended LLM parameters
DIG governance results depend heavily on the underlying large language model (LLM). The following starting values are intended for data governance workloads and most production environments. Always follow the context window, output limit, and supported parameter range of the selected model service.
| Parameter | Recommended value | Guidance |
|---|---|---|
| Model name | DeepSeek-V4-Pro, Kimi K2.6, GLM-5.1, Qwen3.6 | Prefer a model with a long context window (≥128K) and strong Chinese-language support |
| Temperature | 0.1–0.3 | Governance tasks need stable, reproducible output; a high temperature adds randomness |
| Top P | 0.9–1.0 | Retain moderate vocabulary diversity without excessive truncation |
| Max Tokens | 4096–8192 | Keep long prompts and responses from being truncated, especially for batch asset governance |
| Frequency Penalty | 0.0–0.2 | Reduce repeated content; keep the value low |
| Presence Penalty | 0.0–0.2 | Encourage new topics; keep the value low |
| Stop Sequences | No configuration required | Usually leave this unset unless output must end at a specific sequence |
| System Prompt | See the system prompt example | Fix the agent role and output-format constraints |
Hardware and deployment recommendations
For a locally deployed model, use the following as a starting point and load-test it against the model size, concurrency, and context length:
| Resource | Recommended configuration |
|---|---|
| GPU | At least 1 × A100 (40 GB) or 2 × 4090 (24 GB) |
| Memory | ≥ 32 GB |
| Storage | ≥ 200 GB SSD for model cache and logs |
| Network | Low-latency access to the LLM, DAM database, and metadata services |
Model selection guide
The models below are examples organized by governance workload. Confirm current availability, version, and service capabilities with the provider:
| Workload | Suggested models | Rationale |
|---|---|---|
| Metadata completion and standard generation | Qwen 3.7, DeepSeek V4, Kimi 2.5, GLM 5.1 | Strong Chinese semantic understanding and stable output |
| Lineage parsing and code interpretation | Claude 3.5 Sonnet | Strong code understanding for SQL and ETL explanations |
| Classification, grading, and sensitive-data identification | ERNIE 4.0 | Built-in security and compliance semantics |
| Low-cost or test environments | GPT-3.5-Turbo, Qwen-Turbo | Controllable cost for non-critical tasks |
These are general recommendations. Tune them for workload type, data volume, response time, and service cost, then validate the result with an approved small-scope governance task before wider use.
Model types
| Type | Specific settings | Configuration guidance |
|---|---|---|
| LLM | Temperature, maximum tokens | Follow service limits and task requirements; initial form values are not universal recommendations |
| Text Embedding | Vector dimensions | Match the returned vector length; an LLM token limit is not a vector dimension |
Recheck fields after changing type. Before replacing an embedding model or changing dimensions, assess dependent resources and any rebuild requirements.
Fields
| Field | Meaning and verification |
|---|---|
| Configuration name | Identify purpose and service source |
| Model type | LLM or Text Embedding; controls specific fields |
| Provider | Protocol supported by the service |
| Base URL | Actual API root; verify path and reachability |
| API key | Optional in the form, but required when the service requires authentication |
| Discovered model name | Candidate returned by discovery; select the intended model |
| Manual model name | Exact server identifier when discovery is unavailable |
| Default model | System default model; separate from the default agent |
| Temperature | LLM sampling parameter within the service's supported range |
| Maximum tokens | LLM length parameter subject to service limits |
| Vector dimensions | Actual length of the embedding vector |
Keep credentials in authentication fields rather than prompts, descriptions, screenshots, or feedback.
Edit, copy, or delete
The list provides Edit, Copy, and Delete. Check consumers before editing. Retest after changing address, model, or authentication. Rename copied configurations to distinguish their purpose and verify environment-specific values.
Before deletion, confirm that no agent or resource still needs the model. Validate shared changes on an approved small task before wider use.
Validate governance use
- Verify the selected model in agent management.
- Start a test session with a known, authorized object and a bounded query or proposal task.
- Inspect responses, tool calls, and errors.
- Review proposals through approvals and verify DAM results.
- If model testing passes but tool calls fail, check the DAM URL, identity, and dependencies in initial setup.
Troubleshooting
| Symptom | Check |
|---|---|
| Discovery returns no models | Base URL and network, protocol, discovery support, then manual identifier |
| Authentication fails | Whether an API key is needed, valid, and authorized for the target model |
| Cannot save | Connection test and form validation requirements |
| Embedding result mismatch | Actual vector length, configured dimensions, and model type |
| Conversation uses another model | Actual agent and session configuration; distinguish default model from default agent |
| Text works but governance has no result | Tools, baseUrl, identity parameters, and approval status |