find-image

import { createConfigSchematics } from "@lmstudio/sdk";
import { homedir } from "os";
import path from "path";
import { FIND_IMAGE_BACKENDS_ROOT, FIND_IMAGE_MODELS_ROOT, findImageDataPath } from "./paths.js";

/** Sentinel value for multimodalMaxNewEmbeddingsPerRun: at or above this, the run is unlimited (same convention as retrievalLimit's "25 = all"). */
export const MULTIMODAL_MAX_NEW_EMBEDDINGS_UNLIMITED = 100;

export const defaultPluginSettings = {
  // ── Search Settings (per-chat) ────────────────────────────────────────────
  allowByteIdenticalResults: false,
  retrievalLimit: 5,
  gpsToleranceMeters: 0,
  createdToleranceMinutes: 0,
  multimodalMaxNewEmbeddingsPerRun: 25,
  // ── Multimodal Search (Qwen3-VL-Embedding) ───────────────────────────────
  multimodalSearchEnabled: true,
  multimodalEmbeddingBackend: "integrated-local-gguf",
  llamaServerTTL: 0,
  multimodalEmbeddingServicePort: 8765,
  multimodalEmbeddingLmStudioUrl: "http://127.0.0.1:1234",
  multimodalEmbeddingModel: "qwen/qwen3-vl-embedding-8b",
  multimodalEmbeddingModelPath: path.join(
    FIND_IMAGE_MODELS_ROOT,
    "Qwen3-VL-Embedding-8B-GGUF",
  ),
  multimodalEmbeddingGgufBinaryPath: path.join(
    FIND_IMAGE_BACKENDS_ROOT,
    "qwen3-vl-embedding",
    "llama-server"
  ),
  multimodalEmbeddingDimension: 1024,
  multimodalEmbeddingContextSize: 8192,
  multimodalEmbeddingGgufGpuLayers: 999,
  multimodalEmbeddingDataStorePath: findImageDataPath("multimodal_embeddings.sqlite3"),
  embeddingMinScore: 30,
  embeddingDisplayMultiplier: 1.1,
  rerankingEnabled: false,
  rerankingModelPath: path.join(
    FIND_IMAGE_MODELS_ROOT,
    "Qwen3-VL-Reranker-8B-GGUF",
  ),
  rerankingServicePort: 8766,
  rerankingContextSize: 8192,
  rerankingBatchSize: 8192,
  rerankingUbatchSize: 8192,
  rerankingDisplayMultiplier: 1.2,
  rerankingMinScore: 0.05,
  rerankingCandidateMultiplier: 2,
  rerankingMaxCandidates: 25,
  multimodalFlushEvery: 5,
  // ── Search Sources ───────────────────────────────────────────────────────────
  usePreviewsForQueryEmbedding: false,
  contentDirectories: [] as string[],
  searchChatAttachments: true,
  searchWorkingDirectories: true,
  searchDrawThingsProjects: true,
  drawThingsProjectsDirectory: path.join(
    homedir(),
    "Library",
    "Containers",
    "com.liuliu.draw-things",
    "Data",
    "Documents",
  ),
  // ── Cleanup Tools ─────────────────────────────────────────────────────────
  overwriteToken: "",
} as const;

/**
 * Per-chat config. Unlike globalConfigSchematics (application-wide, read via
 * ctl.getGlobalPluginConfig()), changes to these fields take effect
 * immediately for the current conversation (read via ctl.getPluginConfig()).
 * Only search-tuning knobs belong here — nothing that reflects a global
 * resource (model paths, ports, directories, credentials).
 */
export const configSchematics = createConfigSchematics()
  // ── Search Settings ───────────────────────────────────────────────────────
  .field(
    "allowByteIdenticalResults",
    "boolean",
    {
      displayName: "Allow byte-identical Results",
      subtitle: "Show every matching file, including byte-identical copies. When off, only the newest duplicate is shown.",
      engineDoesNotSupport: false,
    },
    defaultPluginSettings.allowByteIdenticalResults,
  )
  .field(
    "retrievalLimit",
    "numeric",
    {
      int: true,
      min: 1,
      max: 25,
      displayName: "Max Results",
      subtitle: "Maximum number of generations (25 = all)",
      slider: { min: 1, max: 25, step: 1 },
    },
    defaultPluginSettings.retrievalLimit,
  )
  .field(
    "gpsToleranceMeters",
    "numeric",
    {
      int: true,
      min: 0,
      max: 100_000,
      displayName: "GPS Tolerance",
      subtitle: "Search radius in metres around the reference GPS coordinates (0 = exact coordinates).",
      slider: { min: 0, max: 5_000, step: 50 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.gpsToleranceMeters,
  )
  .field(
    "createdToleranceMinutes",
    "numeric",
    {
      int: true,
      min: 0,
      max: 43_200,
      displayName: "Created Tolerance",
      subtitle: "Symmetric +/- time window in minutes around the reference creation time (0 = exact time).",
      slider: { min: 0, max: 1_440, step: 1 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.createdToleranceMinutes,
  )
  .field(
    "multimodalMaxNewEmbeddingsPerRun",
    "numeric",
    {
      int: true,
      min: 1,
      max: MULTIMODAL_MAX_NEW_EMBEDDINGS_UNLIMITED,
      displayName: "Max New Multimodal Embeddings per Run",
      subtitle: `Maximum newly generated image embeddings per index run (0 = disabled, ${MULTIMODAL_MAX_NEW_EMBEDDINGS_UNLIMITED} = unlimited)`,
      slider: { min: 0, max: MULTIMODAL_MAX_NEW_EMBEDDINGS_UNLIMITED, step: 1 },
      engineDoesNotSupport: false,
    },
    defaultPluginSettings.multimodalMaxNewEmbeddingsPerRun,
  )
  .build();

export const globalConfigSchematics = createConfigSchematics()
// ── Multimodal Search (Qwen3-VL-Embedding) ───────────────────────────────
  .field(
    "multimodalSearchEnabled",
    "boolean",
    {
      displayName: "Multimodal Search",
      subtitle: "Use Qwen3-VL-Embedding for text-to-image and image-to-image retrieval. When off, no images are indexed/embedded and no query (including image-target queries) uses multimodal search — metadata/keyword search only.",
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.multimodalSearchEnabled,
  )
  .field(
    "multimodalEmbeddingBackend",
    "select",
    {
      displayName: "Multimodal Embedding Backend",
      subtitle: "Backend for Qwen multimodal embeddings. Integrated Local GGUF uses a native llama-server process.",
      options: [
        { value: "integrated-local-gguf", displayName: "Integrated Local GGUF" },
        { value: "lmstudio-api", displayName: "LM Studio API" },
      ],
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.multimodalEmbeddingBackend,
  )
  .field(
    "llamaServerTTL",
    "numeric",
    {
      int: true,
      min: 0,
      max: 120,
      displayName: "llama-server Idle Timeout",
      subtitle: "Minutes to keep the local llama-server running after its last activity. 0 stops it immediately.",
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.llamaServerTTL,
  )
  .field(
    "multimodalEmbeddingServicePort",
    "numeric",
    {
      int: true,
      min: 1,
      max: 65535,
      displayName: "Multimodal Service Port",
      subtitle: "Port for the local llama-server embedding process. Always bound to localhost.",
      engineDoesNotSupport: false,
    },
    defaultPluginSettings.multimodalEmbeddingServicePort,
  )
  .field(
    "multimodalEmbeddingLmStudioUrl",
    "string",
    {
      displayName: "Multimodal LM Studio URL",
      subtitle: "Preferred future LM Studio API URL for multimodal embeddings.",
      placeholder: "http://127.0.0.1:1234",
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.multimodalEmbeddingLmStudioUrl,
  )
  .field(
    "multimodalEmbeddingModel",
    "string",
    {
      displayName: "Multimodal Embedding Model",
      subtitle: "Model identity stored with Qwen multimodal vectors.",
      placeholder: "qwen/qwen3-vl-embedding-8b",
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.multimodalEmbeddingModel,
  )
  .field(
    "multimodalEmbeddingModelPath",
    "string",
    {
      displayName: "Multimodal Model Path",
      subtitle: "Local GGUF model path: an unambiguous directory or an explicit main .gguf file, with a matching mmproj .gguf file alongside it.",
      placeholder: path.join(homedir(), "Documents", "Models", "Qwen3-VL-Embedding-8B-GGUF"),
      engineDoesNotSupport: false,
    },
    defaultPluginSettings.multimodalEmbeddingModelPath,
  )
  .field(
    "multimodalEmbeddingGgufBinaryPath",
    "string",
    {
      displayName: "GGUF Embedding Binary Path",
      subtitle: "Absolute path to a llama-server executable built from the qwen3-vl-embedding llama.cpp fork.",
      placeholder: path.join(homedir(), "Documents", "Repos", "qwen3-vl-embedding", "llama.cpp", "build", "bin", "llama-server"),
      engineDoesNotSupport: false,
    },
    defaultPluginSettings.multimodalEmbeddingGgufBinaryPath,
  )
  .field(
    "multimodalEmbeddingDimension",
    "numeric",
    {
      int: true,
      min: 64,
      max: 4096,
      displayName: "Multimodal Embedding Dimension",
      subtitle: "Matryoshka output dimension used for Qwen multimodal vectors.",
      slider: { min: 64, max: 4096, step: 64 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.multimodalEmbeddingDimension,
  )
  .field(
    "multimodalEmbeddingContextSize",
    "numeric",
    {
      int: true,
      min: 1024,
      max: 32768,
      displayName: "Multimodal Context Size",
      subtitle: "Context window used by the local multimodal embedding runtime.",
      slider: { min: 1024, max: 32768, step: 1024 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.multimodalEmbeddingContextSize,
  )
  .field(
    "multimodalEmbeddingGgufGpuLayers",
    "numeric",
    {
      int: true,
      min: 1,
      max: 999,
      displayName: "GGUF GPU Layers",
      subtitle: "GPU offload layers passed to llama-server (-ngl). GPU is required; use 999 for full offload with a verified patched binary.",
      slider: { min: 1, max: 999, step: 1 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.multimodalEmbeddingGgufGpuLayers,
  )
  .field(
    "multimodalEmbeddingDataStorePath",
    "string",
    {
      displayName: "Multimodal Embedding Data Store Path",
      subtitle: "SQLite database used to store multimodal embeddings.",
      placeholder: findImageDataPath("multimodal_embeddings.sqlite3"),
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.multimodalEmbeddingDataStorePath,
  )
  .field(
    "embeddingMinScore",
    "numeric",
    {
      int: true,
      min: 0,
      max: 100,
      displayName: "Embedding Minimum Score",
      subtitle: "Minimum raw cosine similarity, expressed as a whole-number percent, for Embedding recall.",
      slider: { min: 0, max: 100, step: 5 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.embeddingMinScore,
  )
  .field(
    "embeddingDisplayMultiplier",
    "numeric",
    {
      min: 0.1,
      max: 10,
      displayName: "Embedding Display Multiplier",
      subtitle: "Scales only the displayed Embedding score; recall thresholds and ordering always use the unscaled value.",
      slider: { min: 0.1, max: 10, step: 0.1 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.embeddingDisplayMultiplier,
  )
  .field(
    "rerankingEnabled",
    "boolean",
    {
      displayName: "Enable Reranking",
      subtitle: "Use the configured Q8_0 reranker after Embedding recall for eligible content queries.",
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.rerankingEnabled,
  )
    .field(
    "rerankingServicePort",
    "numeric",
    {
      int: true,
      min: 1,
      max: 65535,
      displayName: "Reranking Service Port",
      subtitle: "Temporary localhost port for the reranker. It must differ from the embedding service port.",
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.rerankingServicePort,
  )
  .field(
    "rerankingModelPath",
    "string",
    {
      displayName: "Reranking Model Path",
      subtitle: "Local Q8_0 reranker GGUF file or its directory, with its F16 mmproj alongside it.",
      placeholder: path.join(homedir(), "Documents", "Models", "Qwen3-VL-Reranker-8B-Q8_0.gguf"),
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.rerankingModelPath,
  )
  .field(
    "rerankingContextSize",
    "numeric",
    {
      int: true,
      min: 1024,
      max: 32768,
      displayName: "Reranking Context Size",
      subtitle: "Maximum tokens for one reranker query-document pair.",
      slider: { min: 1024, max: 32768, step: 1024 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.rerankingContextSize,
  )
  .field(
    "rerankingBatchSize",
    "numeric",
    {
      int: true,
      min: 512,
      max: 32768,
      displayName: "Reranking Batch Size",
      subtitle: "Physical prompt-processing batch size for the temporary reranker server.",
      slider: { min: 512, max: 32768, step: 512 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.rerankingBatchSize,
  )
  .field(
    "rerankingUbatchSize",
    "numeric",
    {
      int: true,
      min: 512,
      max: 32768,
      displayName: "Reranking Microbatch Size",
      subtitle: "Maximum tokens processed per reranker prompt chunk.",
      slider: { min: 512, max: 32768, step: 512 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.rerankingUbatchSize,
  )
  .field(
    "rerankingDisplayMultiplier",
    "numeric",
    {
      min: 0.1,
      max: 10,
      displayName: "Reranking Display Multiplier",
      subtitle: "Scales only the displayed reranker score; recall and reranker ordering always use the unscaled value.",
      slider: { min: 0.1, max: 10, step: 0.1 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.rerankingDisplayMultiplier,
  )
  .field(
    "rerankingMinScore",
    "numeric",
    {
      min: 0,
      max: 1,
      displayName: "Reranking Minimum Score",
      subtitle: "Minimum raw reranker score required to show a reranked candidate; 0 keeps all candidates for calibration.",
      slider: { min: 0, max: 1, step: 0.01 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.rerankingMinScore,
  )
  .field(
    "rerankingCandidateMultiplier",
    "numeric",
    {
      int: true,
      min: 1,
      max: 20,
      displayName: "Reranking Candidate Multiplier",
      subtitle: "Multiplies Max Results to determine how many Embedding recall candidates the reranker evaluates before applying the final result limit.",
      slider: { min: 1, max: 20, step: 1 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.rerankingCandidateMultiplier,
  )
  .field(
    "rerankingMaxCandidates",
    "numeric",
    {
      int: true,
      min: 1,
      max: 100,
      displayName: "Maximum Reranking Candidates",
      subtitle: "Hard cap for one reranking request.",
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.rerankingMaxCandidates,
  )
  .field(
    "multimodalFlushEvery",
    "numeric",
    {
      int: true,
      min: 1,
      max: 100,
      displayName: "Multimodal Store Flush Interval",
      subtitle: "Persist Qwen image embeddings after this many newly embedded images.",
      slider: { min: 1, max: 25, step: 1 },
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.multimodalFlushEvery,
  )
  .field(
    "usePreviewsForQueryEmbedding",
    "boolean",
    {
      displayName: "Use Previews for Image Retrieval",
      subtitle: "Use standardized preview images instead of high-resolution originals. Speeds up the retrieval, but may affect quality.",
      engineDoesNotSupport: false,
    },
    defaultPluginSettings.usePreviewsForQueryEmbedding,
  )
  // ── Search Source 1: Image Directories ────────────────────────────────────
  .field(
    "contentDirectories",
    "stringArray",
    {
      displayName: "Image Directories",
      subtitle: "Search and visually index image files. Metadata is used when available but not required.",
    },
    [...defaultPluginSettings.contentDirectories],
  )
  // ── Search Source 2: LM Studio Chat Attachments ───────────────────────────
  .field(
    "searchChatAttachments",
    "boolean",
    {
      displayName: "Search Chat Attachments",
      subtitle: "Search and visually index images attached to LM Studio chats (~/.lmstudio/user-files)",
      engineDoesNotSupport: false,
    },
    defaultPluginSettings.searchChatAttachments,
  )
  // ── Search Source 3: LM Studio Working Directories ───────────────────────────
  .field(
    "searchWorkingDirectories",
    "boolean",
    {
      displayName: "Search Working Directories",
      subtitle: "Search and visually index images in LM Studio chat working-directories (~/.lmstudio/working-directories)",
    },
    defaultPluginSettings.searchWorkingDirectories,
  )
  // ── Search Source 4: Draw Things Projects (.sqlite3) ──────────────────────
  .field(
    "searchDrawThingsProjects",
    "boolean",
    {
      displayName: "Search Draw Things Projects",
      subtitle: "Search through Draw Things project files with full generation history.",
    },
    defaultPluginSettings.searchDrawThingsProjects,
  )
  .field(
    "drawThingsProjectsDirectory",
    "string",
    {
      displayName: "Projects Directory",
      subtitle: "Directory containing Draw Things project files.",
    },
    defaultPluginSettings.drawThingsProjectsDirectory,
  )
  // ── Cleanup Tools ─────────────────────────────────────────────────────────
  .field(
    "overwriteToken",
    "string",
    {
      displayName: "Overwrite Token",
      subtitle: "Secret token required to overwrite a project file in-place via clean_project. Leave empty to disable in-place overwrite.",
      engineDoesNotSupport: true,
    },
    defaultPluginSettings.overwriteToken,
  )
  .build();