Architecture and complete service inventory

Runtime components

Component	Runs where	Responsibility	Required for
React single-page application	Static HTTPS host/CDN	Dashboard, KB upload, chat, workflow editor	Browser UI
Express TypeScript API	Persistent Node process behind HTTPS	JWT/OAuth, MySQL metadata, S3 upload, SNS publish, chat dispatch, workflow execution, SNS webhook	Entire web product
MySQL 8+	Managed database or server	Users, ownership, content status, assistants, connectors, messages, workflows, execution state	Web API
MongoDB Atlas with Vector Search	MongoDB Atlas	Embedded chunks, restricted queries, URL trees, optional key/custom-DB records	RAG and ingestion
S3 bucket	AWS, baseline us-east-1	Original uploaded documents and media, signed reference access	Uploaded content
Knowledge-base SNS topic	AWS SNS	ADD_FILE, DELETE_FILE, DELETE_KNOWLEDGEBASE dispatch	Asynchronous ingestion
Knowledge-base Python Lambda	AWS Lambda container image from ECR	Parse documents/text, route media, orchestrate crawl, delete vectors	All KB content types
Scraper Python Lambda	AWS Lambda container image from ECR	Headless browser extraction, links, chunks and embeddings	Website knowledge
Indexing-update SNS topic	AWS SNS	Processing/completion/failure and extraction metadata	Web status synchronization
HTTPS SNS subscriber	Express `/knowledgebase/webhook`	Update MySQL from worker status events	End-to-end UI status
Chat-all-models worker	RunPod serverless GPU	Stella embeddings/retrieval, query rewrite, Mongo vector retrieval, GemFilter and generation	Default chat and ALL vector ingestion
Transcriber worker	RunPod serverless GPU	ASR, optional speaker diarization, transcript chunks/indexing	Audio
Explain-vision worker	RunPod serverless GPU	Image descriptions, sampled video scene descriptions, video speech/indexing	Images/video
Container registry	ECR for Lambda; Docker Hub or compatible registry for RunPod	Store deployable images	Build/release
Compatibility chat Lambda	AWS, optional path	Third-party provider chat and credential validation in current web code	OpenAI/Gemini/Cerebras connectors
Google OAuth application	Google, optional	Google login	Social login only
Provider APIs	Optional external services	Third-party chat	Selected provider only

ECR stores images; it does not run Python workers. No separate Redis, queue worker process, Elasticsearch, Pinecone, or vector microservice is required by the inspected active code. No separate always-on Python HTTP service is needed: Flask files are local wrappers; Lambda and RunPod are the deployment entrypoints.

Data flow

flowchart TD
  UI[React UI] --> API[Express API]
  API <--> SQL[(MySQL metadata and sessions)]
  API --> S3[(S3 originals)]
  API --> INPUT[SNS knowledge-base events]
  INPUT --> KB[KB Lambda]
  KB --> SCRAPE[Scraper Lambda]
  KB --> ASR[RunPod transcriber]
  KB --> VISION[RunPod explain vision]
  KB --> CHAT[RunPod chat and embeddings]
  SCRAPE --> CHAT
  ASR --> CHAT
  VISION --> CHAT
  KB --> MONGO[(MongoDB vectors)]
  SCRAPE --> MONGO
  ASR --> MONGO
  VISION --> MONGO
  CHAT <--> MONGO
  API --> CHAT
  KB --> UPDATES[SNS indexing updates]
  SCRAPE --> UPDATES
  ASR --> UPDATES
  VISION --> UPDATES
  UPDATES --> HOOK[Express SNS webhook]
  HOOK --> SQL

Active and compatibility paths

The owner reports chat-all-models as the major working chat backend. In the web snapshot, llmName === 'qwen' calls RunPod directly; all other providers and key validation call the compatibility chat Lambda. Therefore a default-Qwen installation can avoid that chat Lambda, while a feature-complete installation of this web snapshot cannot assume it is unused. The public RunPod handler does not implement provider dispatch merely because the branch is named chat-all-models.

The owner confirmed React + Express as the production web baseline. Laravel and Next.js are excluded from this service inventory and deployment.

Sources: woiceai-web/backend/src/index.ts woiceai-web/backend/src/services/chatAIService.ts woiceai-ai/woice_kb/src/handler.py woiceai-ai/chat-with-all-models/src/handler.py