Global. One model, one vector space — every workspace embeds queries and documents with these settings. Changing anything here on a running deployment requires re-indexing each affected workspace.

Used only when the chat composer's "Think" button is pressed. Leave any field blank to fall back to the search LLM.

Context size is the maximum number of conversation tokens (summary + messages) each chat call carries. When the conversation occupies the threshold fraction of it, the messages are compacted into an anchored note; the full transcript stays in your browser.

Optional. When enabled, every cloud-model call CKB makes (search LLM, think LLM, embeddings, OpenRouter-channel OCR, LightRAG) is routed through ckb-llmproxy so per-call cost lands directly in CKB's usage table — including LightRAG, which has no client-side cost visibility otherwise. When disabled, calls go straight to OpenRouter and only CKB-side calls report cost (LightRAG spend is then not tracked).

Primary OCR path. Preserves structured OCR, tables, and extracted images when available.

OpenRouter Mistral OCR uses the mistral-ocr engine and is normalized to the same app output: text, tables, images, HTML, then PDF.

Vision-only model. Accepts PNG/JPG; reject PDFs (use Mistral nativo or OpenRouter Mistral OCR for those). Pricing per Clarifai: ~$0.20 / 1M input tokens, ~$0.70 / 1M output tokens.

Removes lines that repeat on a large fraction of OCR'd pages (book titles, chapter headers, page numbers). Off by default. Only applies to scanned uploads — native DOCX/EPUB/text-PDF imports are never touched.

Optional second retrieval channel queried in parallel with Qdrant. When enabled, the chat route POSTs each question to LightRAG's /query endpoint and forwards the returned context to the chat LLM alongside Qdrant hits. The chat LLM remains the answerer; LightRAG only contributes context. If LightRAG is unreachable, chat continues without it (logged, not surfaced).

LightRAG-side LLM & embedding

LightRAG itself uses an LLM (for entity extraction and graph queries) and an embedding model — independent from CKB's chat LLM. Saving this section automatically rewrites lightrag-managed.env and restarts the lightrag container — no SSH needed. The binding selects which API protocol LightRAG speaks; the host + model + API key determine which provider/model are actually called.

Exposes CKB to MCP-compatible agents (Claude Desktop, OpenAI Codex, etc.) over Streamable HTTP, mounted on this same port. Two tools: ckb_query and ckb_upload_document. Restart required after enabling/disabling. Keys are per-user; each user manages their own from their profile.

Issued keys (all users)

Failed uploads can leave intermediate files under /data/uploads and /data/html. The canonical artifacts always live in WebDAV; anything here that does not match an indexed/pending book is debris and can be removed safely.