App

This page documents the Streamlit test/prototype app. The production UI is now the React frontend backed by the backend API. Streamlit remains useful as a local test harness and fallback UI.

For the production UI runbook, see Frontend (React).

Entrypoint

streamlit run src/rag/demo/app.py --server.port 8501

See also: LLM pipeline, Diagram, and Stage memory and follow-up chat.

App demo

See it live

Watch the demo: Watch the demo Request access: Contact form

App defaults (authoritative)

Defaults are loaded from DEFAULT_RETRIEVAL_SETTINGS in rag_core/config.py and are used whenever AUTO_PIPELINE_FROM_CHAT=1.

Parameter	Default
dataset	fixed_size
date	2025-09-14
ingest_tag	comm_fixed_C1_g1_2
level	C1
seed_k	120
fetch_top_n	120
expand_ratio	2.0
expand_limit	800
rerank_mode	dense
prefer	summary
alpha	0.7
sort_by	final
diversify	mmr
mmr_lambda	0.5
max_per_doc	2
min_occurrence	3
top_k	20
graph_required	True
coverage	True
fusion	summary

Reranker provider/model default (CLI layer): hf:BAAI/bge-reranker-base unless overridden via --rerank-spec.

AUTO_PIPELINE_FROM_CHAT behavior

AUTO_PIPELINE_FROM_CHAT is parsed as true when set to 1, true, yes, or on.

Setting	Behavior
AUTO_PIPELINE_FROM_CHAT=1	Retrieval settings are hidden/locked; defaults above are used; first Follow-ups Chat question triggers retrieval + synthesis; "Refresh pipeline" is shown; view is forced to Cards. Deployment uses this mode.
AUTO_PIPELINE_FROM_CHAT=0	Retrieval widgets are visible/editable; "Run retrieval" controls the pipeline; user can override all retrieval parameters; view options are visible.

Modes: local vs API

The app can run retrieval locally or call the API:

• RAG_MODE=local (default): Streamlit invokes the CLI wrapper (rag.demo.triple_retriever_runner) which runs python -m rag.retrieval.triple_retriever rerank.
• RAG_MODE=api: Streamlit posts to RAG_API_URL (default http://api:8000) with optional RAG_API_KEY and expects the same retrieval payload.

Overriding dataset/date/ingest_tag

You can override retrieval parameters in two ways:

1) UI (when AUTO_PIPELINE_FROM_CHAT=0): change the sidebar inputs for dataset/date/ingest_tag and rerun retrieval. 2) API (when RAG_MODE=api): pass parameters in the /query payload; the API merges them with defaults.

Inputs and outputs (summary)

• Inputs: GOLD data (chunks + indices), Neo4j for graph expansion, and LLM provider keys.
• Outputs: ranked chunk table, graph visualizations (Pyvis HTML), CSV/HTML exports (when enabled), and optional latency logs.

Neo4j and database notes

• VPS deployment uses Neo4j Community with a single DB: graph-fixed-size.
• Do not rely on multi-db or graph-semantic for production.