[FEATURE] LLM Model Management Dashbaord

[junaidferoz]

Issue: LLM Integration Layer & Dashboard

Summary

CARE currently has no built-in way to talk to LLMs. Every LLM-based feature requires custom components, and there's no cost tracking or prompt management. This issue covers adding a generic LLM integration layer so users can plug in their own API keys, write prompts, and track usage -- all without touching code.

Problem

• Users can't connect to LLMs (OpenAI, Anthropic, Google) without developer intervention.
• There's no way to track what LLM calls cost or what was sent/received.
• The existing NLP broker adds unnecessary latency for simple LLM API calls.
• Prompts are hardcoded per feature instead of being user-configurable.

Proposed Solution

Backend

• A new LLMService that calls provider APIs directly over HTTP, skipping the NLP broker.
• Four new database tables: api_key, llm_provider, llm_log, and prompt_template.
• AES-256-GCM encryption for stored API keys.
• Full I/O logging of every LLM request for cost tracking and research.
• Seeded provider entries for OpenAI, Anthropic, and Google out of the box.

Frontend

• A unified LLM Dashboard page with API key management, prompt template editor, cost breakdown by provider, and a filterable request log.
• An LLM Providers admin page for enabling/disabling providers and restricting models system-wide.
• Vuex store integration following the same pattern as the existing NLP service.

Prompt Templates

• Users write prompts with {{placeholders}} and can preview or test-run them from the dashboard.
• Templates can be shared system-wide, per study, or per project.

Acceptance Criteria

• Users can add, edit, enable/disable, and delete API keys from the dashboard.
• Users can create, edit, duplicate, and delete prompt templates with parameter placeholders.
• Every LLM call is logged with provider, model, tokens, cost, latency, input, and output.
• The dashboard shows usage stats (total requests, tokens, estimated cost).
• The request log supports filtering by provider, status, and time range, plus CSV export.
• Admins can manage providers and control which models are available.
• API keys are encrypted at rest and masked in the UI.
• LLM calls bypass the NLP broker and go directly to provider APIs.

Out of Scope (for now)

• Model browser / comparison page.
• Automatic input mapping from UI components to prompt template parameters.
• Backend enforcement of sharing scopes (study/project-level access control).
• Workflow step integration with prompt templates.

Related Files

Area	Path
Service	backend/webserver/services/llm.js
Encryption	backend/utils/encryption.js
Models	backend/db/models/api_key.js, llm_provider.js, llm_log.js, prompt_template.js
Migrations	backend/db/migrations/20260331100000 through 20260331100006
Dashboard	frontend/src/components/dashboard/LlmDashboard.vue
Admin page	frontend/src/components/dashboard/LlmProviders.vue
Store	frontend/src/store/modules/service.js
Socket	backend/webserver/sockets/service.js