Settings & Configuration¶
v0.0.5 release: the Provider Credentials tab is hidden — LLM dispatch is disabled
The v0.0.5 release ships with the LLM-driven path disabled at compile time pending internal review. The Provider Credentials tab is hidden entirely in v0.0.5 builds, and the model picker on the General tab is non-functional even though the underlying setting persists. The General tab’s provider, default_model, tier, power_user_roles, tier2_pii_redaction, tier2_redact_hostname, rate_limit_per_hour, daily_spend_cap_usd, and session_tool_call_cap fields are documented for the future release that re-enables the LLM path; in v0.0.5, only enabled, mcp_required, mcp_server_url, and the audit-forwarder fields under Audit & Telemetry are operationally meaningful.
The AI Assistant Settings page is at #/settings/ai-assistant within the LogServ App. Admin-only. Four tabs, each handling a different aspect of the AI Assistant configuration.
The Four Tabs¶
| Tab | Scope |
|---|---|
| General | Org-wide AI Assistant defaults (enable/disable, provider, model, tier, MCP gate, server URL, rate limit, spend cap, Power Users, audit forwarder) |
| Provider Credentials | LLM provider API keys (Anthropic / OpenAI / Azure OpenAI / AWS Bedrock) |
| Splunk MCP | MCP Server bearer token + Audit Forwarder HEC token |
| Audit Log | Read-only browser of every audit event in the _ai_assistant_audit index |
In the Templates-only build variant, the Provider Credentials tab is hidden entirely (no LLM provider needed) and an info banner at the top of the page explains the build mode.
General Tab¶
The General tab is divided into four semantic subsections: Feature, Limits & Quotas, Privacy, Audit & Telemetry.
Feature¶
| Field | Default | Description |
|---|---|---|
enabled |
false |
Master switch. When false, the ✦ AI Assistant button in the nav is hidden and no AI traffic flows. The first time an admin flips this to true, an enable-acceptance modal blocks the save until acknowledged. |
provider |
anthropic |
Active LLM vendor: mock / anthropic / openai / azure_openai / bedrock / ollama (future release). |
default_model |
per-provider | Default model id for the active provider. Per-user model picker in the chat panel can switch within the same provider’s models[]. |
tier |
1 |
Privacy tier 0 / 1 / 2. See Privacy Tiers. Elevating to Tier 2 records a vendor_tier2_elevation audit event. |
mcp_required |
true |
When false, runs MCP-less chat mode (streaming-only, no tool dispatch). |
mcp_server_url |
blank | MCP server endpoint. Blank uses the scheme-relative /en-US/splunkd/__raw/services/mcp. See Splunk MCP Setup. |
power_user_roles |
empty CSV | Comma-separated Splunk roles allowed to see the ✦ Power toggle. See Power Mode. |
Limits & Quotas¶
| Field | Default | Description |
|---|---|---|
rate_limit_per_hour |
30 |
Per-user rolling-1-hour rate limit on free-form prompts. 0 = disabled. Maps to LLM10. Canned prompts are never rate-limited. |
session_tool_call_cap |
100 |
Per-chat-session cap on tool dispatches to prevent infinite tool loops. 0 = disabled. Maps to LLM06. |
daily_spend_cap_usd |
per-org | Cumulative-cost cap on free-form vendor spend per app instance per UTC day. Resets at 00:00 UTC. Maps to LLM10. |
Privacy¶
| Field | Default | Description |
|---|---|---|
tier2_pii_redaction |
true |
When Tier 2 is active, redact known identifier columns (email, user(name), *_ip, mac, account) before sending to the LLM. Stable per-value <redacted-XXXXXXX> tags so cardinality reasoning still works. |
tier2_redact_hostname |
false |
Whether to redact hostname columns under the Tier 2 PII rule. Default off — hostname is often non-sensitive in SAP environments. |
Audit & Telemetry¶
| Field | Default | Description |
|---|---|---|
audit_forwarder_enabled |
false |
Forward audit events to a separate Splunk / SIEM via HEC for tamper-evidence. When the admin saves with this off, a forwarder-disabled-acceptance modal blocks the save until acknowledged. |
audit_forwarder_url |
blank | HEC endpoint URL (e.g., https://siem.example.com:8088/services/collector). |
audit_forwarder_index |
blank | Optional index field to send with each event. |
audit_forwarder_source |
logserv:ai_assistant:audit |
source field for forwarded events. |
Provider Credentials Tab¶
One panel per provider. Each panel has the credential fields for that provider — typically just an API key, plus per-provider extras (Azure deployment URL, Bedrock region, etc.). Credentials are stored in Splunk’s encrypted password store via /servicesNS/nobody/<app>/storage/passwords. The Settings page only ever displays length + prefix, never the cleartext.
The realm convention is logserv_ai_assistant_<provider> and the credential name corresponds to the field. So Anthropic’s API key lives at realm logserv_ai_assistant_anthropic name api_key; OpenAI’s at logserv_ai_assistant_openai name api_key; etc.
Setting a credential (admin-only):
- Pick the provider’s panel on this tab.
- Click Set next to the field.
- Paste the credential.
- Click Save.
- The Settings page records a
local_onlyaudit event with the realm + name + length + prefix (never the cleartext).
Deleting a credential:
- Click Delete next to the field.
- Confirm.
- The credential is removed from Splunk’s encrypted password store.
If the active provider field on the General tab points at a provider whose credentials aren’t set, the AI Assistant produces an error on the first free-form prompt: “Provider ‘X’ has no credentials configured. Open Settings → Provider Credentials to set them.”
In the Templates-only build variant, this entire tab is hidden.
Splunk MCP Tab¶
Two panels:
Splunk MCP Server¶
| Field | Realm | Name |
|---|---|---|
bearer_token |
logserv_ai_assistant_mcp |
bearer_token |
OAuth/JWT token issued by your Splunk MCP Server. The admin pastes this; a future release will mint it automatically. Optional — cookie auth from the same Splunk Web session works by default for HTTP-only Splunk.
Audit Log Forwarder¶
| Field | Realm | Name |
|---|---|---|
hec_token |
logserv_ai_assistant_forwarder |
hec_token |
HEC token for tamper-evident audit forwarding to a separate Splunk / SIEM. The destination URL + on/off toggle live under General → Audit & Telemetry. The token is sent on every audit-event POST as Authorization: Splunk <token>. See Audit Log → HEC Forwarder.
Audit Log Tab¶
Read-only browser of every event in the _ai_assistant_audit index. Filters: time range (preset Last 24h / 7d / 30d / 90d), category multi-select (12 categories — local_only, vendor_tier1, vendor_tier2, security_blocked_spl, rate_limited_prompt, user_prompt_jailbreak_flag, session_tool_cap_hit, daily_spend_cap_hit, audit_forwarder_failure, vendor_tier2_elevation, forwarder_disabled_acceptance, ai_assistant_enable_acceptance), user-contains text filter, result limit (25 / 100 / 500 / 1000).
Clicking a row’s + expand button reveals the full event JSON. The page is paginated client-side at 25 rows per page; “Showing N-M of T events” + Previous / Next buttons render below the table when T > 25.
Tamper-resistance disclaimer at the top of the panel:
Audit events live in a Splunk index. A host-root admin can edit the bucket files. Mitigation: forward audit events to a separate Splunk instance / SIEM / S3-with-Object-Lock owned by a different admin team. Configure the HEC forwarder under Settings → General → Audit & Telemetry.
See Audit Log for the threat model + forwarder configuration.
Legal Acknowledgement Modals¶
Two compile-time legal modals gate specific Settings save flows. Both follow Splunk’s splunk_instrumentation optInVersion framework — once acknowledged at version V, future saves with the same disclaimer revision skip the modal; bumping the version forces re-ack.
Enable Acceptance Modal (orange-bordered)¶
Triggers when:
- saved.enabled === false && draft.enabled === true (every “I’m turning this on” deliberate action), OR
- The current optInVersion of the enable-TC stanza is higher than the user’s last acknowledgement.
The modal disclaimer covers seven clauses: data egress acknowledgement, customer responsibility, AS-IS warranty disclaimer, indemnification, limitation of liability, authority warrant, record-of-acknowledgement. Aligned to Splunk MSA + Cisco EULA conventions.
User identity, Splunk-stamped IP (host field), timestamp, and SHA-256 of the disclaimer revision are recorded in an ai_assistant_enable_acceptance audit event. Yes path: save proceeds + bumps the user’s optInVersionAcknowledged. No path: save aborts + records the No choice.
Forwarder Disabled Acceptance Modal (cyan-bordered)¶
Triggers when:
- saved.audit_forwarder_enabled === true && draft.audit_forwarder_enabled === false (every “I’m turning this protection off” deliberate action), OR
- The current optInVersion of the forwarder-TC stanza is higher than the user’s last acknowledgement.
The disclaimer covers integrity-mitigation responsibility: with the forwarder off, audit events are recoverable only from the local index, which a host-root admin can edit. Same identity + IP + timestamp + disclaimer-hash recording as the enable modal. Records forwarder_disabled_acceptance audit event.
Why the hybrid trigger logic¶
Pure optInVersion-only gating (acknowledge once, never prompted again until the version bumps) is what Splunk’s stock telemetry uses. We added the deliberate-toggle-transition trigger because re-enabling the AI Assistant after a disable, or re-disabling the forwarder after enabling, are deliberate user actions that warrant re-acknowledgement of the legal posture each time. Pure version-only gating would let these toggle-flips slip without re-ack.
Live UI Refresh on Save¶
When the admin saves a config change in the General tab that affects what users see (master enabled toggle, mcp_required, power_user_roles, etc.), the AI Assistant re-applies the new config to the running session within a few seconds. No page reload required.
For example: disabling the AI Assistant via the master toggle and clicking Save Defaults — the ✦ AI Assistant button in the top-right nav disappears within ~4 seconds, no browser refresh needed. Similarly, re-enabling re-shows the button immediately. For the callback wiring, see AI Assistant Implementation Reference.