# agent-project

**Repository Path**: RRRRRDong/agent-project

## Basic Information

- **Project Name**: agent-project
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-06-04
- **Last Updated**: 2026-06-09

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# AIGC Project

React + Ant Design front end with a Python FastAPI LangGraph agent service.

## Stack

- Frontend: Vite, React, TypeScript, Ant Design
- Backend: FastAPI, LangGraph, LangChain OpenAI integration
- Model gateway: NewAPI-compatible OpenAI API

## Implemented scenario

This project now includes the "爆款视频复刻改造" workflow:

- Create a replication project from a competitor viral video URL or uploaded video.
- Add the merchant's own product images or describe the target product in natural language.
- Analyze the viral structure with LangGraph agents.
- Generate product-replacement, cross-border localization, audience, selling-point, style, or scene variants from the same workflow.
- Submit a render job that creates per-shot video tasks through a configurable NewAPI video adapter.

When `NEWAPI_VIDEO_MODE=mock`, render jobs produce JSON manifests under `/generated/...` so the full product flow can be tested without a real video model.

It also includes the "爆款图文批量复刻" workflow:

- Create a content replication project from a viral article, link, or topic.
- Decompose the source into topic, title pattern, body pattern, emotion curve, and story angles.
- Let the user assign custom counts per generated story angle, then generate exactly that many pieces.
- Run large content pools as background jobs, batching generation in small chunks and exposing progress for polling.
- Try to read public HTML from pasted links; if the source cannot be fetched, paste the article text instead.
- Generate copy and image together in the same content job: copy uses `NEWAPI_MODEL` such as `deepseek-v4-pro`, and images use `NEWAPI_IMAGE_MODEL` such as `gpt-image-2`.

When `NEWAPI_IMAGE_MODE=mock`, image tasks produce SVG previews under `/generated/...`. Set `NEWAPI_IMAGE_MODE` to a non-`mock` value and configure `NEWAPI_IMAGE_MODEL`/`NEWAPI_IMAGE_ENDPOINT` to call a real image model such as `gpt-image-2`.

## Agent framework choice

This project uses LangGraph. It gives the Python agent an explicit state graph, so the product can grow from a planner into multi-step analysis, replacement strategy, compliance review, rendering, retry, and memory workflows without changing the service shape.

## Setup

```bash
cp backend/.env.example backend/.env
npm install
npm --prefix frontend install
python3 -m venv backend/.venv
backend/.venv/bin/pip install -r backend/requirements.txt
```

Edit `backend/.env`:

```dotenv
NEWAPI_DOMAIN=https://your-newapi-domain
NEWAPI_API_KEY=sk-...
NEWAPI_MODEL=deepseek-v4-pro
NEWAPI_SCRIPT_MODEL=deepseek-v4-pro
NEWAPI_QA_MODEL=deepseek-v4-pro
NEWAPI_SOURCE_DIALOGUE_MODEL=doubao-seed-2-0-pro-260215
NEWAPI_SOURCE_DIALOGUE_ENDPOINT=/v1/responses
NEWAPI_VIDEO_MODE=mock
NEWAPI_VIDEO_PAYLOAD_MODE=generation
NEWAPI_VIDEO_MODEL=doubao-seedance-2-0-fast-260128
NEWAPI_VIDEO_ENDPOINT=
NEWAPI_VIDEO_RESULT_ENDPOINT=
NEWAPI_IMAGE_MODEL=gpt-image-2
NEWAPI_IMAGE_MODE=mock
NEWAPI_IMAGE_ENDPOINT=/v1/images/generations
NEWAPI_IMAGE_SIZE=1024x1024
STORAGE_DIR=storage
PUBLIC_BASE_URL=
MYSQL_ENABLED=false
MYSQL_HOST=127.0.0.1
MYSQL_PORT=3306
MYSQL_USER=root
MYSQL_PASSWORD=
MYSQL_DATABASE=aigc_project
MYSQL_CHARSET=utf8mb4
```

For a real video model, set `NEWAPI_VIDEO_MODE` to a non-`mock` value and configure:

- `NEWAPI_VIDEO_MODEL`: model name used in the video request payload.
- `NEWAPI_VIDEO_ENDPOINT`: absolute URL or path under `NEWAPI_DOMAIN`.
- `NEWAPI_VIDEO_RESULT_ENDPOINT`: optional polling path, supports `{task_id}` replacement.
- `NEWAPI_VIDEO_PAYLOAD_MODE`: `generation` sends a conservative text-to-video payload for remix/localization workflows; `reference_content` includes source video/product image references. The "仅换产品" mode always uses reference content and requires a video endpoint that supports source-video/product-image references.
- `NEWAPI_SOURCE_DIALOGUE_ENDPOINT`: `/v1/responses` endpoint used by SourceDialogueAgent to extract audio text from a public source-video URL.
- `PUBLIC_BASE_URL`: public base URL for `/generated/...` assets when the video model needs to fetch uploaded source videos or product images.

Optional local MySQL mirror for generated content data:

```sql
CREATE DATABASE IF NOT EXISTS aigc_project
  DEFAULT CHARACTER SET utf8mb4
  DEFAULT COLLATE utf8mb4_unicode_ci;
```

Then set `MYSQL_ENABLED=true` in `backend/.env`. The backend keeps the existing JSON files as the primary local cache and mirrors content replication data into MySQL tables:

- `content_projects`: latest project snapshot and history list fields.
- `content_jobs`: latest background generation job snapshot.
- `backend_data_events`: append-only project/job data event log.
- `backend_api_logs`: request/response snapshots for `/api/*` and `/health` routes.

Run both services:

```bash
npm run dev
```

- Frontend: http://localhost:5173
- Backend: http://localhost:8100 (`backend/app` Python file changes hot reload automatically)
- API docs: http://localhost:8100/docs