Imported from RainLib/FerrisMind (.agent/skills/project-guidelines/SKILL.md). Install upstream with npx skills add RainLib/FerrisMind --skill project-guidelines. Copyright stays with the author.

FerrisMind / openNotebookLm Project Guidelines

You are working on the FerrisMind project, an open-source AI knowledge workspace heavily inspired by NotebookLM but with a unique, premium design identity and a Rust-driven backend.

Whenever you write or modify code for this project, you MUST adhere to the following architecture, styling, and structural guidelines.

1. Tech Stack

• Frontend: Next.js (App Router), React, TypeScript.
• Styling: Tailwind CSS v4 (configured purely via @theme in src/app/globals.css, there is no tailwind.config.ts), vanilla CSS.
• Backend (API): Rust, Axum, Async-GraphQL.
• Database: SurrealDB (native graph DB — uses RELATE for graph edges, -> / <- for graph traversal).
• AI/LLM: Multi-model abstraction via rig-core (Gemini, OpenAI, Anthropic, DeepSeek). Embedding via same framework.
• Search: Hybrid Search — Vector (cosine similarity on chunk embeddings) + Knowledge Graph (entity/relation graph traversal).

2. UI / UX Design System (Amber Tech-Chic Linear)

The project utilizes a highly distinct, precision-engineered "Tech-Chic" aesthetic. Do not use generic, default Tailwind components. You must follow these visual rules:

Colors & Palette

• Backgrounds: White (bg-white inside workspace), Off-white/Stone-50 (bg-bg-sources, bg-bg-studio for sidebars).
• Foreground/Borders: Softer Black (#171717, used as border-black, text-black or border-border-bold).
• Accents: Amber (accent-main for primary actions like borders/buttons, accent-secondary for hovers, accent-light for backgrounds/highlights).
• Secondary Text: Gray (text-gray-400, text-gray-500 for subtitles and metadata).

Shadows & Components

• NEVER use soft blurry shadows! The design relies entirely on hard, offset linear shadows.
• Use the custom CSS variables defined in globals:
  • shadow-hard (3px 3px 0px 0px #171717)
  • shadow-hard-sm (1px 1px 0px 0px #171717)
  • shadow-modal (for floating modals, combination of soft drop and hard offset)
• For hover states, components should shift upwards hover:-translate-y-0.5 or hover:-translate-y-1 and increase shadow depth (e.g., hover:shadow-hard-hover).
• Borders: Elements must use 1px solid borders (border border-black or border border-gray-200/300). Use dashed borders for drop zones or empty states.
• Corners: Keep corners sharp or very slightly rounded (rounded-sm or rounded-none). No pill shapes (rounded-full) except for specific toggles or avatars.

Typography

• Primary Font: Public Sans or Inter.
• Headings: font-black, tracking-tight, uppercase.
• Labels/Overheads: Extremely small, spaced-out uppercase tags must be heavily used: text-[10px] font-bold uppercase tracking-widest text-gray-400/500.
• Tooltips: Use absolute positioned, black background (bg-black), white text (text-white), tiny font (text-[10px] font-bold) tooltips instead of native HTML title attributes whenever possible.

Icons

• Use Google Material Symbols Outlined.
• Structure: <span className="material-symbols-outlined icon-sm">icon_name</span>
• Available sizing classes: text-[16px], icon-sm (18px), or icon-lg (24px).

Background Patterns

• Utilize the custom background stripes/hatches defined in globals (bg-background-image-diagonal-hatch, bg-background-image-diagonal-pattern) for empty states, drop zones, or decorative headers.

3. Frontend Architecture

• Editor Layout: The primary app lies in src/components/editor/EditorLayout.tsx. It uses a 3-pane structure: LeftSidebar (Sources), ChatPanel (Center), and RightSidebar (Studio Tools).
• Responsiveness: The desktop layout uses react-resizable-panels for drag-to-resize columns. The mobile layout (isMobile) uses absolute full-screen overlays with swipe/toggle logic (CollapsedLeftSidebar, CollapsedRightSidebar).
• State Management: Sidebar toggle state, width dragging, and selection logic are handled locally using React hooks. When building complex components, encapsulate state into the feature component rather than polluting the global layout.

4. Backend Architecture

Ingest Pipeline (src/ingest/)

• Pipeline: IngestPipeline — parse → sanitize → chunk → embed (via EmbeddingProvider).
• Service: IngestionService in service.rs — orchestrates streaming ingestion, stores chunks with embeddings, tracks chunk_id_map (chunk_index → DB record id).
• KG Extraction: KgExtractor in kg_extractor.rs:
  • After all chunks are embedded and stored, the full article text is assembled and split into token-bounded windows using tiktoken-rs (cl100k_base, default 8K tokens/window).
  • Each window → LLM → strict JSON schema response with entities and relations.
  • Entities stored as kg_entity nodes; relations via RELATE entity_a->kg_relation->entity_b.
  • KG extraction is async/background (non-blocking to main ingest flow).

Knowledge Graph Data Model (SurrealDB)

• kg_entity — Node table: notebook, document, chunk_id, label, entity_type, properties, is_active, created_at.
• kg_relation — Edge table (TYPE RELATION IN kg_entity OUT kg_entity): notebook, relation_type, confidence, chunk_id, is_active, created_at. SurrealDB manages in/out automatically.
• Soft-delete: Removing a source → is_active = false on its KG records (never hard-delete). All KG queries filter by is_active = true.
• Notebook isolation: All KG records carry a notebook field, subgraphs are per-notebook.

Query Flows (src/graph/)

All query flows use graph-flow crate (FlowRunner/GraphBuilder) to compose task DAGs with shared ChatFlowData context.

• Chat flow (chat.rs): ChatContextTask → ChatKgSearchTask → ChatSearchTask → ChatResponseTask
• Ask flow (ask.rs): AskEntryTask → AskSearchTask → AskKgSearchTask → AskQueryProcessTask → AskFinalAnswerTask
• KG Search (kg_search.rs): KgSearcher — single SurrealQL query does label fuzzy-match + 1-hop bidirectional graph expansion via -> and <- operators.
• Context fields: ChatFlowData carries kg_hits, kg_context, search_results, sub_answers etc.

LLM Manager (src/llm/)

• LlmManager wraps RigClient + PromptManager.
• llm.agent() for no-preamble, llm.agent_with_preamble(...) for custom preambles.
• AnyAgent supports .prompt(), .prompt_with_retry(), .stream_to_sse().

Database Schema (src/db/schema.rs)

• Applied at startup via idempotent DEFINE statements.
• Tables: user, notebook, has_access (RELATION), document, doc_image, chunk, session, message, kg_entity, kg_relation (RELATION).

5. Workflows & Rules

1. Modifying UI: Refer to Shadows & Components and Typography guidelines. Match the "Tech-Chic" vibe. No generic blue buttons or soft shadows.
2. Adding Dependencies: Ask before adding new npm/Cargo packages.
3. Icons Check: Use the exact string name for material-symbols-outlined icons.
4. Tailwind Config: Modify values in src/app/globals.css via @theme, not tailwind.config.ts.
5. KG Conventions:
   • Extracted KG data must carry notebook + document references.
   • Removing sources → soft-delete KG (is_active = false), never hard-delete.
   • Query flows must include kg_context in prompt template variables.
6. Graph Flows: Each query type is a task DAG in src/graph/. Adding a stage = create a Task struct → add to GraphBuilder → wire edges.

Whenever taking an action on the UI, confirm it aligns with the strict visual hierarchy and Amber/Off-black aesthetic.