Predefined embedding modes in search

.sqlx/query-d56ec6cf262f4f17cd2aaa1a6413bf0fec1d7a35661cfd02feae01e94b807b7b.json

@@ -0,0 +1,26 @@
+{
+  "db_name": "SQLite",
+  "query": "SELECT * FROM predefined_embeddings",
+  "describe": {
+    "columns": [
+      {
+        "name": "name",
+        "ordinal": 0,
+        "type_info": "Text"
+      },
+      {
+        "name": "embedding",
+        "ordinal": 1,
+        "type_info": "Blob"
+      }
+    ],
+    "parameters": {
+      "Right": 0
+    },
+    "nullable": [
+      false,
+      false
+    ]
+  },
+  "hash": "d56ec6cf262f4f17cd2aaa1a6413bf0fec1d7a35661cfd02feae01e94b807b7b"
+}

clipfront2/src/App.svelte

@@ -60,6 +60,8 @@
                 align-items: center
                 > *
                     margin: 0 2px
+        .sliders-ctrl
+            width: 5em
 
     .result
         border: 1px solid gray
@@ -70,6 +72,9 @@
 </style>
 
 <h1>Meme Search Engine</h1>
+{#if config.n_total}
+    <p>{config.n_total} items indexed.</p>
+{/if}
 <details>
     <summary>Usage tips</summary>
     <ul>
@@ -78,6 +83,7 @@
         <li>In certain circumstances, it may be useful to postfix your query with "meme".</li>
         <li>Capitalization is ignored.</li>
         <li>Only English is supported. Other languages might work slightly.</li>
+        <li>Sliders are generated from PCA on the index. The human-readable labels are approximate.</li>
     </ul>
 </details>
 <div class="controls">
@@ -98,12 +104,21 @@
                 {#if term.type === "embedding"}
                     <span>[embedding loaded from URL]</span>
                 {/if}
+                {#if term.type === "predefined_embedding"}
+                    <span>{term.predefinedEmbedding}</span>
+                {/if}
             </li>
         {/each}
     </ul>
     <div class="ctrlbar">
         <input type="search" placeholder="Text Query" on:keydown={handleKey} on:focus={newTextQuery}>
         <button on:click={pickFile}>Image Query</button>
+        <select bind:value={predefinedEmbeddingName} on:change={setPredefinedEmbedding} class="sliders-ctrl">
+            <option>Sliders</option>
+            {#each config.predefined_embedding_names ?? [] as name}
+                <option>{name}</option>
+            {/each}
+        </select>
         <button on:click={runSearch} style="margin-left: auto">Search</button>
     </div>
 </div>
@@ -151,6 +166,25 @@
     let queryTerms = []
     let queryCounter = 0
 
+    let config = {}
+    util.serverConfig.subscribe(x => {
+        config = x
+    })
+    let predefinedEmbeddingName = "Sliders"
+
+    const setPredefinedEmbedding = () => {
+        if (predefinedEmbeddingName !== "Sliders") {
+            queryTerms.push({
+                type: "predefined_embedding",
+                predefinedEmbedding: predefinedEmbeddingName,
+                sign: "+",
+                weight: 0.2
+            })
+        }
+        queryTerms = queryTerms
+        predefinedEmbeddingName = "Sliders"
+    }
+
     const decodeFloat16 = uint16 => {
         const sign = (uint16 & 0x8000) ? -1 : 1
         const exponent = (uint16 & 0x7C00) >> 10
@@ -200,7 +234,7 @@
     let displayedResults = []
     const runSearch = async () => {
         if (!resultPromise) {
-            let args = {"terms": queryTerms.map(x => ({ image: x.imageData, text: x.text, embedding: x.embedding, weight: x.weight * { "+": 1, "-": -1 }[x.sign] }))}
+            let args = {"terms": queryTerms.filter(x => x.text !== "").map(x => ({ image: x.imageData, text: x.text, embedding: x.embedding, predefined_embedding: x.predefinedEmbedding, weight: x.weight * { "+": 1, "-": -1 }[x.sign] }))}
             queryCounter += 1
             resultPromise = util.doQuery(args).then(res => {
                 error = null

clipfront2/src/util.js

@@ -1,4 +1,5 @@
 import * as config from "../../frontend_config.json"
+import { writable } from "svelte/store"
 
 export const getURL = x => config.image_path + x[1]
 
@@ -16,4 +17,10 @@ export const hasFormat = (results, result, format) => {
 
 export const thumbnailURL = (results, result, format) => {
     return `${config.thumb_path}${result[2]}${format}.${results.extensions[format]}`    
-}
+}
+
+export let serverConfig = writable({})
+fetch(config.backend_url).then(x => x.json().then(x => {
+    serverConfig.set(x)
+    window.serverConfig = x
+}))

load_embedding.py

@@ -0,0 +1,14 @@
+import numpy as np
+import sqlite3
+import base64
+
+#db = sqlite3.connect("/srv/mse/data.sqlite3")
+db = sqlite3.connect("data.sqlite3")
+db.row_factory = sqlite3.Row
+
+name = input("Name: ")
+url = input("Embedding search URL: ")
+data = base64.urlsafe_b64decode(url.removeprefix("https://mse.osmarks.net/?e="))
+arr = np.frombuffer(data, dtype=np.float16).copy()
+db.execute("INSERT OR REPLACE INTO predefined_embeddings VALUES (?, ?)", (name, arr.tobytes()))
+db.commit()

src/main.rs

@@ -27,6 +27,7 @@ use tower_http::cors::CorsLayer;
 use faiss::index::scalar_quantizer;
 use lazy_static::lazy_static;
 use prometheus::{register_int_counter, register_int_counter_vec, register_int_gauge, Encoder, IntCounter, IntGauge, IntCounterVec};
+use ndarray::ArrayBase;
 
 mod ocr;
 
@@ -93,6 +94,11 @@ CREATE VIRTUAL TABLE IF NOT EXISTS ocr_fts USING fts5 (
     content='files'
 );
 
+CREATE TABLE IF NOT EXISTS predefined_embeddings (
+    name TEXT NOT NULL PRIMARY KEY,
+    embedding BLOB NOT NULL
+);
+
 CREATE TRIGGER IF NOT EXISTS ocr_fts_ins AFTER INSERT ON files BEGIN
     INSERT INTO ocr_fts (rowid, filename, ocr) VALUES (new.rowid, new.filename, COALESCE(new.ocr, ''));
 END;
@@ -127,10 +133,11 @@ struct InferenceServerConfig {
     embedding_size: usize,
 }
 
-#[derive(Debug, Deserialize, Clone)]
+#[derive(Debug, Clone)]
 struct WConfig {
     backend: InferenceServerConfig,
-    service: Config
+    service: Config,
+    predefined_embeddings: HashMap<String, ArrayBase<ndarray::OwnedRepr<f32>, ndarray::prelude::Dim<[usize; 1]>>>
 }
 
 async fn query_clip_server<I, O>(
@@ -699,6 +706,7 @@ struct QueryTerm {
     embedding: Option<EmbeddingVector>,
     image: Option<String>,
     text: Option<String>,
+    predefined_embedding: Option<String>,
     weight: Option<f32>,
 }
 
@@ -760,6 +768,10 @@ async fn handle_request(config: Arc<WConfig>, client: Arc<Client>, index: &IInde
                 total_embedding[i] += value * weight;
             }
         }
+        if let Some(name) = &term.predefined_embedding {
+            let embedding = config.predefined_embeddings.get(name).context("name invalid")?;
+            total_embedding = total_embedding + embedding * term.weight.unwrap_or(1.0);
+        }
     }
     
     let mut batches = vec![];    
@@ -806,6 +818,12 @@ async fn get_backend_config(config: &Config) -> Result<InferenceServerConfig> {
     Ok(rmp_serde::from_slice(&res.bytes().await?)?)
 }
 
+#[derive(Serialize, Deserialize)]
+struct FrontendInit {
+    n_total: u64,
+    predefined_embedding_names: Vec<String>
+}
+
 #[tokio::main]
 async fn main() -> Result<()> {
     pretty_env_logger::init();
@@ -826,9 +844,21 @@ async fn main() -> Result<()> {
         }
     };
 
+    let mut predefined_embeddings = HashMap::new();
+
+    {
+        let db = initialize_database(&config).await?;
+        let result = sqlx::query!("SELECT * FROM predefined_embeddings")
+            .fetch_all(&db).await?;
+        for row in result {
+            predefined_embeddings.insert(row.name, ndarray::Array::from(decode_fp16_buffer(&row.embedding)));
+        }
+    }
+
     let config = Arc::new(WConfig {
         service: config,
-        backend
+        backend,
+        predefined_embeddings
     });
 
     if config.service.no_run_server {
@@ -878,6 +908,8 @@ async fn main() -> Result<()> {
 
     let config_ = config.clone();
     let client = Arc::new(Client::new());
+    let index_ = index.clone();
+    let config__ = config.clone();
     let app = Router::new()
         .route("/", post(|req| async move {
             let config = config.clone();
@@ -886,10 +918,13 @@ async fn main() -> Result<()> {
             QUERIES_COUNTER.inc();
             handle_request(config, client.clone(), &index, req).await.map_err(|e| format!("{:?}", e))
         }))
-        .route("/", get(|_req: axum::http::Request<axum::body::Body>| async move {
-            "OK"
+        .route("/", get(|_req: ()| async move {
+            Json(FrontendInit {
+                n_total: index_.read().await.vectors.ntotal(),
+                predefined_embedding_names: config__.predefined_embeddings.keys().cloned().collect()
+            })
         }))
-        .route("/reload", post(|_req: axum::http::Request<axum::body::Body>| async move {
+        .route("/reload", post(|_req: ()| async move {
             log::info!("Requesting index reload");
             let mut done_rx = done_tx.clone().subscribe();
             let _ = request_ingest_tx.send(()).await; // ignore possible error, which is presumably because the queue is full
@@ -911,7 +946,7 @@ async fn main() -> Result<()> {
                 }
             }
         }))
-        .route("/metrics", get(|_req: axum::http::Request<axum::body::Body>| async move {
+        .route("/metrics", get(|_req: ()| async move {
             let mut buffer = Vec::new();
             let encoder = prometheus::TextEncoder::new();
             let metric_families = prometheus::gather();