Adaptive · Corrective · Self-RAG

Upload documents or paste text, tune chunking, and build the dense vectorstore. This is the private source the adaptive router can route to — the live web is the corrective fallback when a query's retrieved passages are graded off-topic.

Upload documents

Drag & drop files here, or browse

.txt · .md · .pdf

…or paste text directly

Label for pasted text Content

Chunking

Chunk size 800 chars Chunk overlap 120 chars

Embedding model: — · configured in Settings

No index yet — add content above, then click Build / Rebuild Index. You can still chat: the router will send questions to the live web or answer directly.

Retrieval, loop guards & display

KB passages per search 4 Web results per search 5 Min relevant to skip fallback 1 Max query rewrites 2 loop guard Max regenerations 2 loop guard Allow web-search correction Show graph trajectory Show sources

Settings are locked

Enter the settings password to view or change the model, embedding & web-search configuration.

Configuration

Stored server-side in SQLite (app_config.db) — never in the browser.

Model—

Region—

Bedrock key—

Web search—

Model & connection

AWS region Bedrock model ID Bedrock API key Bedrock base URL override optional Embedding model (local, sentence-transformers)

Changing the embedding model requires rebuilding the index. anthropic.* models get fully-streamed generation & tool use; other families (e.g. qwen.*) use the OpenAI-compatible endpoint.

Live web search (corrective fallback)

DuckDuckGo DuckDuckGo needs no key. Add a Tavily key for cleaner, content-rich results tuned for RAG.

Tavily API key optional

Security

Current password New password min 6 chars Confirm new password