71 lines
3.3 KiB
Markdown
71 lines
3.3 KiB
Markdown
# Implementation Plan - Captain V2 Multi-LLM & Documentation
|
|
|
|
This plan addresses three key objectives:
|
|
|
|
1. **Correcting the Context:** Updating `.context/docs` to reflect our specific WuzAPI/Jasmine/Captain architecture.
|
|
2. **Multi-LLM Architecture:** Implementing a flexible LLM provider system (OpenAI/Gemini) for Captain.
|
|
3. **Unification:** Integrating Jasmine's brain (Intent/SDR logic) into Captain's body (AgentRunner).
|
|
|
|
## User Review Required
|
|
|
|
> [!IMPORTANT] > **Database Schema Change**: We will add `llm_provider` and `llm_model` columns to `captain_inbox_configs`.
|
|
> **Configuration Strategy**: This allows _per-inbox_ configuration. One inbox can use GPT-4o (Sales), another Gemini 1.5 Flash (Support).
|
|
|
|
## Proposed Changes
|
|
|
|
### 1. Correcting the Context (Documentation)
|
|
|
|
We will create a specific architecture addendum file to avoid overwriting the generic generic Chatwoot docs every time.
|
|
|
|
#### [NEW] [.context/docs/custom-architecture.md](file:///Users/user/Chatwoot/chatwoot-develop/.context/docs/custom-architecture.md)
|
|
|
|
- **Content:**
|
|
- **WuzAPI Integration Module:** Explains the V1 Contract (Encrypted Tokens, Webhook Handlers, Sync Logic).
|
|
- **Jasmine Brain:** Explains Intent Detection, Strategy Decider, and Playbooks.
|
|
- **Captain V2 (The Body):** Explains `AgentRunnerService` and how it delegates to the Brain.
|
|
|
|
### 2. Multi-LLM Backend Infrastructure
|
|
|
|
#### [MODIFY] [db/migrate/timestamp_add_llm_config_to_captain.rb](file:///Users/user/Chatwoot/chatwoot-develop/db/migrate/20260101000000_add_llm_config_to_captain.rb)
|
|
|
|
- Add columns to `captain_inbox_configs` (or `chat_assistants` table, checking existing schema first):
|
|
- `llm_provider` (string, default: 'openai')
|
|
- `llm_model` (string, default: 'gpt-3.5-turbo')
|
|
- `api_key` (text, encrypted)
|
|
|
|
#### [MODIFY] [enterprise/app/services/captain/llm/provider_factory.rb](file:///Users/user/Chatwoot/chatwoot-develop/enterprise/app/services/captain/llm/provider_factory.rb)
|
|
|
|
- **New Service:** A factory pattern to instantiate the correct client.
|
|
- `def self.get_client(provider, config)`
|
|
- Case `openai`: Returns `OpenAI::Client`
|
|
- Case `gemini`: Returns `Gemini::QuickClient` (from our definition)
|
|
|
|
#### [MODIFY] [enterprise/app/services/captain/assistant/agent_runner_service.rb](file:///Users/user/Chatwoot/chatwoot-develop/enterprise/app/services/captain/assistant/agent_runner_service.rb)
|
|
|
|
- Update `generate_response` to use `ProviderFactory` instead of hardcoded OpenAI logic.
|
|
|
|
### 3. Jasmine Logic Injection
|
|
|
|
#### [MODIFY] [enterprise/app/services/captain/llm/jasmine_brain.rb](file:///Users/user/Chatwoot/chatwoot-develop/enterprise/app/services/captain/llm/jasmine_brain.rb)
|
|
|
|
- Ensure this class implements the "Thinking" step:
|
|
1. Detect Intent (Sales/Support/Spam).
|
|
2. Select Strategy.
|
|
3. Assemble Prompt.
|
|
|
|
## Verification Plan
|
|
|
|
### Automated Tests
|
|
|
|
- **Migration:** Run `rails db:migrate` and verify schema.
|
|
- **Provider Factory:** Create a Rails Console script to request a completion from Gemini and OpenAI via the factory.
|
|
|
|
### Manual Verification
|
|
|
|
1. **Configure:** Set Inbox 1 to use **Gemini 1.5 Flash**.
|
|
2. **Chat:** Send a message "Quero comprar um carro".
|
|
3. **Verify:**
|
|
- Logs should show `[Captain] Using provider: gemini`
|
|
- Response should come from Gemini.
|
|
- Intent should be logged as `sales`.
|