Cognosa Component Stack

Runtime processes, the packages inside them, the libraries they are built on, and the stores they talk to

The platform is three runtime pieces plus a set of offline tools. The React single-page app runs in the browser. The app process serves it and the REST API. The rt process executes the RAG pipeline. All Python code shares the common package. LangChain lives only in rt and the ingestion tools: it is the retrieval and chat abstraction, never part of the web API.

Columns are the runtime processes; the panels inside each are the repo packages in call order, top to bottom, with the third-party libraries at the foot. The shared common package spans everything Python. Arrows show which process opens a connection to which store or endpoint. Only Qdrant is drawn as the vector store; ChromaDB and pgvector implement the same vdb_ops interface.

Browser app process rt process Offline tools Shared package Data store Third-party libraries

Where LangChain is used

Package	Used in	For
`langchain_core`	`tasks_lib/llm_lib/llm_type_*.py`	PromptTemplate, RunnablePassthrough, RunnableLambda, StrOutputParser; the chain is `{context, question} \| prompt \| llm \| parser`, run with `.stream()`
`langchain_openai`	`llm_type_openai.py`	ChatOpenAI for OpenAI, Qwen (DashScope), and Ollama local or remote through the OpenAI-compatible endpoint
`langchain_anthropic`, `langchain_google_genai`	`llm_type_claude.py`, `llm_type_gemini.py`	ChatAnthropic, ChatGoogleGenerativeAI
`langchain_huggingface` + sentence-transformers	`vdb_lib/emb_models.py`, loaders	HuggingFaceEmbeddings, preloaded per model name from `RT_VDB_EMB_MODELS_PRELOAD`
`langchain_qdrant`, `langchain_chroma`, `langchain_community` PGVector	`vdb_lib/*_ops.py`	Vector store wrappers; `as_retriever(search_type, search_kwargs)` with a Qdrant payload filter built from the group's retrieval filters
`langchain_community` loaders, `langchain_text_splitters`	`document_loader_*.py`, `tools/ingest_dev/`	PDF, DOCX, PPTX, XLSX, TXT loading and chunking at ingestion time

Runtime processes

Process	Entry point	Concurrency	Talks to
React SPA	`frontend/src/main.tsx`, built by Vite into `backend/static`	Browser; polls `GET /api/v1/doc_tasks/{uuid}` until the task finishes	app only, over HTTPS with a JWT bearer
app	`backend/webapp.py` under uvicorn	Async; one worker	PostgreSQL (async SQLAlchemy); reads VDB and LLM health from `api_processes`
rt	`backend/run_tasks.py`	Main loop; `RT_VDB_PROCESS_NUM` VDB worker processes; one LLM thread per task; watchdog threads	PostgreSQL (sync SQLAlchemy), Qdrant, Ollama hosts, cloud LLM APIs, Hugging Face Hub on first model load
Offline tools	Run by hand from `backend/`	Loaders use a multiprocessing pool per file	Qdrant (ingestion), PostgreSQL (init, migrate, backup)