┌─────────────┐     ┌──────────────┐     ┌─────────────────┐
│  Axum HTTP  │────▶│   Pipeline   │────▶│  Gemini API      │
│  :8000      │     │  Orchestrator│     │  embed + generate │
└─────────────┘     └──────┬───────┘     └─────────────────┘
                           │
          ┌────────────────┼────────────────┐
          ▼                ▼                ▼
   ┌──────────┐    ┌──────────┐     ┌──────────┐
   │ Storage  │    │Extraction│     │ Qdrant   │
   │ local/R2 │    │ pdf/docx │     │ Vector   │
   │          │    │ xlsx/pptx│     │ Store    │
   └──────────┘    │ image/txt│     └──────────┘
                   └──────────┘

src/
├── main.rs              — Startup: model discovery, storage prompt, server bootstrap
├── config.rs            — All env-var configuration
├── error.rs             — thiserror-based AppError + Result<T>
├── server/
│   ├── mod.rs           — Router: POST /api/v1/rag/query
│   ├── handlers.rs      — Request handler with auth + pipeline orchestration
│   └── auth.rs          — Bearer token extraction
├── pipeline/
│   ├── mod.rs           — Pipeline: process → embed → search → answer
│   ├── download.rs      — download_bytes(url) → Vec<u8>
│   └── url.rs           — extract_filename_from_url(url) → String
├── extraction/
│   ├── mod.rs           — TextExtractor trait
│   ├── pdf.rs           — pdf-extract + lopdf (pure Rust)
│   ├── docx.rs          — zip + quick-xml (pure Rust)
│   ├── xlsx.rs          — calamine (pure Rust)
│   ├── pptx.rs          — zip + quick-xml (pure Rust)
│   ├── image.rs         — image crate → OcrEngine
│   ├── text.rs          — fs::read_to_string
│   └── libreoffice.rs   — Optional soffice fallback
├── ocr/
│   ├── mod.rs           — OcrEngine trait
│   └── paddle.rs        — ocrs + rten (RTen-based OCR, auto-downloads models)
├── ai/
│   ├── mod.rs           — Module re-exports
│   ├── traits.rs        — EmbeddingProvider + LlmProvider traits
│   └── gemini/
│       ├── mod.rs       — GeminiProvider: model discovery + constructor
│       ├── client.rs    — reqwest Client builder, backoff/retry helpers
│       ├── embed.rs     — EmbedClient: impl EmbeddingProvider
│       ├── llm.rs       — LlmClient: impl LlmProvider
│       ├── types.rs     — All serde structs + ModelInfo
│       ├── safety.rs    — sanitize_policy() prompt injection defense
│       └── models.rs    — discover_models() API + interactive selection
├── vectordb/
│   ├── mod.rs           — VectorStore trait + ChunkEmbedding/ScoredChunk
│   └── qdrant.rs        — QdrantStore: full gRPC CRUD + cosine search
└── storage/
    ├── mod.rs           — StoredFile struct + StorageBackend trait
    ├── local.rs         — LocalStorage: files on disk
    └── r2.rs            — R2Storage: Cloudflare R2 via aws-sdk-s3

URL
  ↓  download_bytes(url)
Vec<u8>
  ↓  StoredFile::new(filename, len)
StoredFile { id: uuid, storage_key, mime_type }
  ↓  storage.put(key, bytes, mime)
  ↓  storage.get_local_path(key) → PathBuf
  ↓  extractor.extract_text(&Path) → String
  ↓  chunk_text(text, 8000 chars)
  ↓  embed_provider.embed(chunk) → Vec<f32>  × N chunks
  ↓  vector_store.store_embeddings(doc_id, chunks)
  ↓  embed_provider.embed(questions) → query vector
  ↓  vector_store.search_similar(query, top_k, threshold)
  ↓  llm_provider.generate(context + questions, schema) → JSON answers
[Qdrant: cosine similarity, 3072-dim vectors]

POST /api/v1/rag/query
Authorization: Bearer <token>
Content-Type: application/json

{
    "documents": "https://example.com/document.pdf",
    "questions": [
        "What is the grace period?",
        "What does section 4.1 cover?"
    ]
}

→ {
    "answers": [
        "The grace period is 30 days...",
        "Section 4.1 covers..."
    ]
}