@ -16,15 +16,15 @@ Ghostfolio is a privacy-first, open-source personal finance dashboard for tracki
### What AgentForge Adds (The AI Layer)
AgentForge extends Ghostfolio's **existing AiModule** (not a new module) to add a full conversational agent:
AgentForge adds a **new, independent `AgentModule`** — separate from the existing `AiModule` — providing a full conversational agent:
- **LangGraph TS ReAct agent** — `createReactAgent` with tool-calling loop and verification post-node
- **LangGraph TS StateGraph** — 6-node pipeline: `query → plan → execute → verify → disclaim → respond`
- **Model-agnostic LLM provider** — defaults to Claude Sonnet 4 via `@langchain/anthropic` , swappable to any provider
- **7 tools** wrapping existing Ghostfolio services (portfolio summary, performance, holdings, activities, market data, risk analysis, account overview)
- **Real-time streaming** — Server-Sent Events (SSE) for streaming AI responses via new `POST /api/v1/ai/chat` endpoint
- **Conversation memory** — persistent chat history stored in PostgreSQL via Prisma (`Conversation` + `Message` models )
- **Real-time streaming** — Server-Sent Events (SSE) for streaming AI responses via `POST /api/v1/agents/chat`
- **Conversation memory** — LangGraph checkpointer persisted to PostgreSQL (no custom Prisma models needed )
- **LangSmith observability** — tracing, latency metrics, and tool-call auditing
- **Admin-configurable LLM** — API keys and model selection stored via PropertyService (already exists)
- **Extended Assistant UI** — Angular Assistant component gains a chat tab alongside existing search
- **Minimal chat UI** — dedicated `/chat` page with message history and streaming display
---
@ -37,7 +37,7 @@ AgentForge extends Ghostfolio's **existing AiModule** (not a new module) to add
| **Frontend** | Angular 21.x with Angular Material |
| **Database** | PostgreSQL 15 via Prisma 6.x ORM |
| **Cache** | Redis (Bull queues for background jobs) |
| **AI/LLM** | LangGraph TS (`@langchain/langgraph`), Vercel AI SDK , LangSmith |
| **AI/LLM** | LangGraph TS (`@langchain/langgraph`), `@langchain/anthropic`
| **Auth** | JWT, Google OAuth, OIDC, WebAuthn (Passport strategies) |
| **Containerization** | Docker / Docker Compose |
| **i18n** | Angular i18n (12 languages) |
@ -226,84 +226,87 @@ Key models in `prisma/schema.prisma`:
---
## Existing AI Infrastructure
## Existing AI Infrastructure (Reference Only)
Ghostfolio already has a lightweight AI module. AgentForge extends it rather than replacing it.
### Current State
Ghostfolio has an existing lightweight `AiModule` (from PR #4176 ). AgentForge does **not** extend it — we build a new, independent `AgentModule` . The existing module is documented here for reference only.
| Component | Location | What It Does |
|-----------|----------|-------------|
| **AiController** | `apps/api/src/app/endpoints/ai/ai.controller.ts` | `GET /api/v1/ai/prompt/:mode` — returns a formatted prompt string |
| **AiService** | `apps/api/src/app/endpoints/ai/ai.service.ts` | `getPrompt()` builds a markdown holdings table; `generateText()` calls OpenRouter via Vercel AI SDK |
| **AiModule** | `apps/api/src/app/endpoints/ai/ai.module.ts` | Imports PortfolioService, AccountService, MarketDataService, etc. |
| **Assistant UI** | `libs/ui/src/lib/assistant/assistant.component.ts` | Search/navigation modal (accounts, holdings, asset profiles, quick links) — **not yet a chat UI** |
| **PropertyService** | `apps/api/src/services/property/property.service.ts` | Stores `API_KEY_OPENROUTER` and `OPENROUTER_MODEL` in the `Property` DB table |
| **Config constants** | `libs/common/src/lib/config.ts` | `PROPERTY_API_KEY_OPENROUTER` , `PROPERTY_OPENROUTER_MODEL` , `PROPERTY_SYSTEM_MESSAGE` |
| **Permission** | `libs/common/src/lib/permissions.ts` | `readAiPrompt` — granted to ADMIN, DEMO, USER roles |
| **AI SDK deps** | `package.json` | `ai` (Vercel AI SDK 4.x), `@openrouter/ai-sdk-provider` |
### Current Data Flow
| `AiController` | `apps/api/src/app/endpoints/ai/ai.controller.ts` | `GET /api/v1/ai/prompt/:mode` — returns a formatted prompt string |
| `AiService` | `apps/api/src/app/endpoints/ai/ai.service.ts` | `getPrompt()` builds a markdown holdings table; `generateText()` calls OpenRouter via Vercel AI SDK |
| `AiModule` | `apps/api/src/app/endpoints/ai/ai.module.ts` | Imports PortfolioService, AccountService, MarketDataService, etc. |
| `Assistant UI` | `libs/ui/src/lib/assistant/assistant.component.ts` | Search/navigation modal — **not a chat UI** |
1. User opens Analysis page → clicks "Copy AI Prompt" (portfolio or analysis mode)
2. Frontend calls `GET /api/v1/ai/prompt/{mode}` with optional filters
3. `AiService.getPrompt()` fetches holdings via `PortfolioService.getDetails()` , formats a markdown table
4. Prompt is copied to clipboard; user pastes into external LLM (Duck.ai, ChatGPT, etc.)
**Key insight:** The current flow is prompt-generation-only with no in-app chat, no tool calling, and no conversation memory.
**We do not touch any of these files.** Our `AgentModule` is fully independent.
---
## AgentForge Integration Plan (v3 )
## AgentForge Integration Plan (v2)
The core principle: **extend the existing `AiModule`** — do not create a new module. Add files alongside `ai.controller.ts` and `ai.service.ts` .
### Core Principle
### Architecture: LangGraph TS ReAct Agent
**New, independent `AgentModule` ** — does not modify or depend on the existing `AiModule` . Clean separation, own endpoints, own state management.
A single `createReactAgent` (from `@langchain/langgraph` ) with a tool-calling loop and a **verification post-node** that reviews tool outputs before responding to the user.
### Architecture: LangGraph TS StateGraph (6 Nodes)
```
User message
│
▼
┌─────────┐ ┌───────────┐ ┌──────────────┐
│ Agent │────▶│ Tools │────▶│ Verification │
│ (ReAct) │◀────│ (7 tools) │◀────│ Post-Node │
└─────────┘ └───────────┘ └──────────────┘
│
▼
Streamed response (SSE)
┌─────────┐ ┌────────┐ ┌───────────┐ ┌──────────┐ ┌───────────┐ ┌───────────┐
│ Query │──▶│ Plan │──▶│ Execute │──▶│ Verify │──▶│ Disclaim │──▶│ Respond │
│ (parse) │ │(decide)│ │ (tools) │ │ (checks) │ │(caveats) │ │ (stream) │
└─────────┘ └────────┘ └───────────┘ └──────────┘ └───────────┘ └───────────┘
```
### Extended Module Structure
| Node | Purpose |
|------|---------|
| **Query** | Parse user input, classify intent, extract entities |
| **Plan** | Decide which tools to call and in what order |
| **Execute** | Run tools via LangGraph tool-calling loop (may cycle back to Plan) |
| **Verify** | Domain-specific validation — sanity-check numbers, flag stale data, validate ranges |
| **Disclaim** | Inject financial disclaimers ("not financial advice", data freshness caveats) |
| **Respond** | Format final response and stream via SSE |
### Module Structure
```
apps/api/src/app/endpoints/ai/
├── ai.module.ts # Extended with new providers
├── ai.controller.ts # Extended with POST /chat and GET /conversations
├── ai.service.ts # Extended with agent orchestration
├── tools/ # NEW — LangGraph tool definitions
│ ├── portfolio-summary.tool.ts # Wraps PortfolioService.getDetails()
│ ├── portfolio-performance.tool.ts # Wraps PortfolioService.getPerformance()
│ ├── holdings-lookup.tool.ts # Wraps PortfolioService.getHoldings()
│ ├── activity-search.tool.ts # Wraps OrderService
│ ├── market-data.tool.ts # Wraps DataProviderService
│ ├── risk-analysis.tool.ts # Wraps RulesService / portfolio rules
│ └── account-overview.tool.ts # Wraps AccountService
├── memory/ # NEW — Conversation persistence
│ └── conversation.service.ts # CRUD for Conversation + Message via Prisma
└── streaming/ # NEW — SSE streaming
└── sse.service.ts # Server-Sent Events for token delivery
apps/api/src/app/endpoints/agents/
├── agent.module.ts # NestJS module (independent from AiModule)
├── agent.controller.ts # REST + SSE endpoints
├── agent.service.ts # Core orchestration — builds and invokes the graph
├── graph/
│ ├── agent.graph.ts # StateGraph definition with 6 nodes
│ ├── state.ts # AgentState type definition
│ └── nodes/
│ ├── query.node.ts # Intent classification + entity extraction
│ ├── plan.node.ts # Tool selection + execution planning
│ ├── execute.node.ts # Tool invocation via LangGraph
│ ├── verify.node.ts # Domain-specific output validation
│ ├── disclaim.node.ts # Financial disclaimer injection
│ └── respond.node.ts # Response formatting + SSE streaming
├── tools/
│ ├── portfolio-summary.tool.ts # Wraps PortfolioService.getDetails()
│ ├── portfolio-performance.tool.ts # Wraps PortfolioService.getPerformance()
│ ├── holdings-detail.tool.ts # Wraps PortfolioService.getHoldings()
│ ├── activity-history.tool.ts # Wraps OrderService
│ ├── market-data.tool.ts # Wraps DataProviderService
│ ├── account-overview.tool.ts # Wraps AccountService
│ └── risk-report.tool.ts # Wraps RulesService + portfolio rules
├── providers/
│ └── llm-provider.ts # Model-agnostic LLM provider interface
└── streaming/
└── sse.service.ts # Server-Sent Events for token delivery
```
### New API Endpoints
### API Endpoints
| Method | Route | Purpose |
|--------|-------|---------|
| `POST` | `/api/v1/ai/chat` | Send message, get streamed SSE response |
| `GET` | `/api/v1/ai/conversations` | List user's conversations |
| `GET` | `/api/v1/ai/conversations/:id` | Get conversation with messages |
| `DELETE` | `/api/v1/ai/conversations/:id` | Delete a conversation |
| `POST` | `/api/v1/agents /chat` | Send message, get streamed SSE response |
| `GET` | `/api/v1/agents /conversations` | List user's conversations |
| `GET` | `/api/v1/agents /conversations/:id` | Get conversation with messages |
| `DELETE` | `/api/v1/agents /conversations/:id` | Delete a conversation |
### 7 Tools (LangGraph `DynamicStructuredTool` )
@ -311,48 +314,43 @@ Each tool wraps an existing Ghostfolio service — **no duplicate business logic
| Tool Name | Wraps | Input Schema | Returns |
|-----------|-------|-------------|---------|
| `portfolio_summary` | `PortfolioService.getDetails()` | filters (accounts, tags, assetClasses) | Holdings table with allocations, sectors, currencies |
| `portfolio_performance` | `PortfolioService.getPerformance()` | dateRange, filters | ROI, TWR, MWR, chart data for time range |
| `holdings_lookup ` | `PortfolioService.getHoldings()` | symbol (optional), filters | Detailed holding info (quantity, price, P& L) |
| `activity_search ` | `OrderService.getOrders()` | symbol, type, dateRange | Filtered transaction history |
| `portfolio_summary` | `PortfolioService.getDetails()` | filters (accounts, tags, assetClasses) | Holdings with allocations, sectors, currencies |
| `portfolio_performance` | `PortfolioService.getPerformance()` | dateRange, filters | ROI, TWR, MWR, chart data |
| `holdings_detai l` | `PortfolioService.getHoldings()` | symbol (optional), filters | Quantity, price, P& L per holding |
| `activity_history ` | `OrderService.getOrders()` | symbol, type, dateRange | Filtered transaction history |
| `market_data` | `DataProviderService.getQuotes()` | symbols[] | Current quotes, daily change, market state |
| `risk_analysis ` | `RulesService` + portfolio rules | filters | Rule evaluations (cluster risk, currency risk, etc.) |
| `risk_report ` | `RulesService` + portfolio rules | filters | Rule evaluations (cluster risk, currency risk, etc.) |
| `account_overview` | `AccountService.getAccounts()` | accountId (optional) | Account balances, platforms, cash positions |
### New Prisma Models
```prisma
model Conversation {
id String @id @default (uuid())
createdAt DateTime @default (now())
messages Message[]
title String?
updatedAt DateTime @updatedAt
user User @relation (fields: [userId], references: [id], onDelete: Cascade)
userId String
}
model Message {
id String @id @default (uuid())
content String
conversation Conversation @relation (fields: [conversationId], references: [id], onDelete: Cascade)
conversationId String
createdAt DateTime @default (now())
role String // 'user' | 'assistant' | 'tool'
toolCalls Json? // Serialized tool invocations
tokenCount Int?
}
### Conversation Memory
Uses **LangGraph checkpointer** persisted to PostgreSQL — no custom Prisma `Conversation` /`Message` models needed. LangGraph manages state serialization and retrieval natively via `@langchain/langgraph-checkpoint-postgres` .
Each conversation is identified by a `thread_id` (UUID). The checkpointer stores the full graph state (messages, tool results, metadata) at each step.
### LLM Provider
Model-agnostic provider interface with **Claude Sonnet 4** as default:
```typescript
// providers/llm-provider.ts
import { ChatAnthropic } from '@langchain/anthropic';
// Default: Claude Sonnet 4 via Anthropic API
// Swappable to OpenAI, OpenRouter, or any @langchain/ * provider
```
### Frontend: Extended Assistant Component
API key stored via `PropertyService` (key: `API_KEY_ANTHROPIC` ). Model configurable via `ANTHROPIC_MODEL` property.
The existing `GfAssistantComponent` (`libs/ui/src/lib/assistant/`) gains a **chat tab** alongside the current search functionality:
### Frontend: Dedicated Chat Page
- New tab or mode toggle: **Search** (existing) | **Chat** (new)
- Chat tab renders conversation history with streaming message display
- Input field sends messages to `POST /api/v1/ai/chat`
- SSE consumption for real-time token rendering
- Conversation list sidebar for switching between chats
New lazy-loaded Angular route at `/chat` :
- Full-page layout with conversation list sidebar + chat area
- Streaming message display via SSE (`EventSource`)
- Input field sends to `POST /api/v1/agents/chat`
- Basic markdown rendering for agent responses
- Shows tool call indicators (e.g., "Fetching portfolio summary...")
### Observability: LangSmith
@ -363,13 +361,33 @@ The existing `GfAssistantComponent` (`libs/ui/src/lib/assistant/`) gains a **cha
### Key Design Principles
1. **Extend, don't fork** — add to the existing `AiModule` ; keep all current prompt-generation functionality working
1. **Independent module** — `AgentModule` is fully separate from `AiModule` ; no shared state, no shared code
2. **Wrap, don't replace** — tools call existing services; no duplicate business logic
3. **Verification post-node** — LangGraph graph includes a node after tool execution that validates outputs before responding
4. **Streaming-first** — all chat responses use SSE for real-time token delivery
5. **Permission-gated** — tool access respects existing `readAiPrompt` permission and user roles
6. **Admin-configurable** — LLM provider, model, and API keys stored in PropertyService (no env vars needed)
7. **Auditable** — LangSmith tracing for all tool invocations; Message model stores toolCalls JSON
3. **6-node pipeline** — query → plan → execute → verify → disclaim → respond
4. **Verify node** — domain-specific validation (sanity-check portfolio values, flag stale data, validate date ranges)
5. **Disclaim node** — automatic financial disclaimers on all responses
6. **Streaming-first** — all chat responses use SSE for real-time token delivery
7. **Model-agnostic** — swap LLM provider without code changes; just update PropertyService
8. **LangGraph-native memory** — checkpointer handles conversation persistence, no custom ORM models
9. **Auditable** — LangSmith tracing for all tool invocations
---
## 24-Hour MVP Requirements (Hard Gate)
All items required to pass:
| # | Requirement | Implementation |
|---|------------|----------------|
| 1 | Agent responds to natural language queries | LangGraph StateGraph with Claude Sonnet 4 |
| 2 | At least 3 functional tools | `portfolio_summary` , `holdings_detail` , `account_overview` |
| 3 | Tool calls execute successfully with structured results | Tools wrap existing services, return typed JSON |
| 4 | Agent synthesizes tool results into coherent responses | Respond node formats tool outputs into natural language |
| 5 | Conversation history maintained across turns | LangGraph checkpointer to PostgreSQL |
| 6 | Basic error handling (graceful failure, not crashes) | Try/catch in each node, fallback error messages |
| 7 | At least one domain-specific verification check | Verify node: validate portfolio totals, flag stale market data |
| 8 | 5+ test cases with expected outcomes | Jest test suite in `agent.service.spec.ts` |
| 9 | Deployed and publicly accessible | Docker build via existing `docker-compose.build.yml` |
---
Claude
1 month ago