update chunked embed

env

@@ -9,7 +9,8 @@ ROUTE_INTERNAL=true
 DOWNLOAD_ON_DEMAND=true
 
 # Blacklisted domains (Comma-seperated regex)
-BLACKLIST_DOMAINS="google.com,.*.youtube.com"
+# You can blacklist sites which wont work well
+BLACKLIST_DOMAINS="^gitlab"
 
 # Database
 DATABASE_URL=postgres://user:pass@postgres/webarc

migrations/0001_embedding.sql

@@ -5,6 +5,7 @@ CREATE TABLE doc_embedding (
     domain VARCHAR(500) NOT NULL,
     path VARCHAR(1000) NOT NULL,
     ver VARCHAR(10) NOT NULL,
+    chunk INTEGER NOT NULL,
     embed_mxbai_embed_large vector(1024) NOT NULL,
-    PRIMARY KEY (domain, path, ver)
+    PRIMARY KEY (domain, path, ver, chunk)
 )

src/ai.rs

@@ -1,4 +1,4 @@
-use std::collections::VecDeque;
+use std::collections::{HashMap, VecDeque};
 
 use based::{get_pg, request::api::ToAPI, result::LogNoneAndPass};
 use ollama_rs::generation::embeddings::request::{EmbeddingsInput, GenerateEmbeddingsRequest};
@@ -8,11 +8,15 @@ use sqlx::FromRow;
 
 use crate::archive::{Document, Domain, WebsiteArchive};
 
+// TODO : Chunked embeddings + better search + ranking
+// TODO : Real citese embeddings + search
+
 #[derive(Debug, Clone, FromRow, Serialize)]
 pub struct DocEmbedding {
     pub domain: String,
     pub path: String,
     pub ver: String,
+    pub chunk: i32,
 
     #[serde(skip)]
     embed_mxbai_embed_large: pgvector::Vector,
@@ -21,6 +25,53 @@ pub struct DocEmbedding {
     pub similarity: f64,
 }
 
+#[derive(Debug, Clone, Serialize)]
+pub struct SearchResult {
+    pub domain: String,
+    pub path: String,
+    pub chunks: Vec<DocEmbedding>,
+}
+
+impl SearchResult {
+    pub fn new(domain: String, path: String) -> Self {
+        Self {
+            domain,
+            path,
+            chunks: vec![],
+        }
+    }
+
+    pub fn similarity(&self) -> f64 {
+        total_score(&self.chunks)
+    }
+}
+
+pub fn avg_sim(e: &[DocEmbedding]) -> f64 {
+    let mut score = 0.0;
+
+    for e in e {
+        score += e.similarity;
+    }
+
+    score / e.len() as f64
+}
+
+pub fn max_sim(e: &[DocEmbedding]) -> f64 {
+    let mut score = 0.0;
+
+    for e in e {
+        if e.similarity > score {
+            score = e.similarity;
+        }
+    }
+
+    score
+}
+
+pub fn total_score(e: &[DocEmbedding]) -> f64 {
+    (avg_sim(e) + max_sim(e)) / 2.0
+}
+
 impl ToAPI for DocEmbedding {
     async fn api(&self) -> serde_json::Value {
         json!({
@@ -33,12 +84,23 @@ impl ToAPI for DocEmbedding {
 }
 
 pub trait Embedding {
-    fn embedding(&self, ver: Option<String>)
-        -> impl std::future::Future<Output = Option<Vec<f32>>>;
+    fn embedding(
+        &self,
+        ver: Option<String>,
+    ) -> impl std::future::Future<Output = Option<Vec<Vec<f32>>>>;
+}
+
+pub fn chunked(s: &str) -> Vec<String> {
+    const CHUNK_SIZE: usize = 500;
+    s.chars()
+        .collect::<Vec<char>>()
+        .chunks(CHUNK_SIZE)
+        .map(|chunk| chunk.iter().collect())
+        .collect()
 }
 
 impl Embedding for Document {
-    async fn embedding(&self, ver: Option<String>) -> Option<Vec<f32>> {
+    async fn embedding(&self, ver: Option<String>) -> Option<Vec<Vec<f32>>> {
         let latest = "latest".to_string();
         log::info!(
             "Generating Vector embeddings for {} / {} @ {}",
@@ -47,9 +109,26 @@ impl Embedding for Document {
             ver.as_ref().unwrap_or(&latest)
         );
 
-        let content_html = self.render_local(ver).await?;
+        let content_html = self.render_local(ver.clone()).await?;
         let content = html2md::parse_html(&content_html);
-        generate_embedding(content).await
+
+        let mut embeddings = Vec::new();
+        let content = chunked(&content);
+        let len = content.len();
+
+        for (index, c) in content.into_iter().enumerate() {
+            log::info!(
+                "Generating Vector embeddings for {} / {} @ {} [ {} / {} ]",
+                self.domain,
+                self.path,
+                ver.as_ref().unwrap_or(&latest),
+                index + 1,
+                len
+            );
+            embeddings.push(generate_embedding(c).await?);
+        }
+
+        Some(embeddings)
     }
 }
 
@@ -129,14 +208,17 @@ impl EmbedStore {
             .execute(get_pg!())
             .await;
 
-            sqlx::query("INSERT INTO doc_embedding VALUES ($1, $2, $3, $4)")
-                .bind(&doc.domain)
-                .bind(&doc.path)
-                .bind(ver)
-                .bind(embed)
-                .execute(get_pg!())
-                .await
-                .unwrap();
+            for (index, embed) in embed.iter().enumerate() {
+                sqlx::query("INSERT INTO doc_embedding VALUES ($1, $2, $3, $4, $5)")
+                    .bind(&doc.domain)
+                    .bind(&doc.path)
+                    .bind(ver)
+                    .bind(index as i64)
+                    .bind(embed)
+                    .execute(get_pg!())
+                    .await
+                    .unwrap();
+            }
         }
     }
 
@@ -148,8 +230,12 @@ impl EmbedStore {
         }
     }
 
-    pub async fn search_vector(v: &pgvector::Vector, limit: i64, offset: i64) -> Vec<DocEmbedding> {
-        sqlx::query_as(
+    pub async fn search_vector(v: &pgvector::Vector, limit: i64, offset: i64) -> Vec<SearchResult> {
+        // TODO : fix search
+        // + new ranked algorithm
+        // + better repr
+
+        let results: Vec<DocEmbedding> = sqlx::query_as(
             "SELECT *, 1 / (1 + (embed_mxbai_embed_large <-> $1)) AS similarity FROM doc_embedding ORDER BY embed_mxbai_embed_large <-> $1 LIMIT $2 OFFSET $3",
         )
         .bind(v)
@@ -157,7 +243,28 @@ impl EmbedStore {
         .bind(offset)
         .fetch_all(get_pg!())
         .await
-        .unwrap()
+        .unwrap();
+
+        let mut search_res: HashMap<String, HashMap<String, SearchResult>> = HashMap::new();
+
+        for res in results {
+            let domain = search_res
+                .entry(res.domain.clone())
+                .or_insert(HashMap::new());
+            let doc = domain
+                .entry(res.path.clone())
+                .or_insert(SearchResult::new(res.domain.clone(), res.path.clone()));
+            doc.chunks.push(res);
+        }
+
+        let mut flat = search_res
+            .into_values()
+            .map(|x| x.into_values().collect::<Vec<SearchResult>>())
+            .flatten()
+            .collect::<Vec<SearchResult>>();
+
+        flat.sort_by(|a, b| b.chunks.len().cmp(&a.chunks.len()));
+        flat
     }
 
     pub async fn generate_embeddings_for(arc: &WebsiteArchive) {

src/archive.rs

@@ -1,6 +1,6 @@
 use std::{io::Read, path::PathBuf};
 
-use based::request::RequestContext;
+use based::{request::RequestContext, result::LogAndIgnore};
 use maud::html;
 
 use crate::{blacklist::check_blacklist, favicon::download_fav_for, pages::component::render_page};
@@ -62,7 +62,8 @@ impl Domain {
     /// A new `Domain` instance.
     pub fn new(name: &str, dir: PathBuf) -> Self {
         if !check_blacklist(name) {
-            std::fs::create_dir_all(&dir).unwrap();
+            std::fs::create_dir_all(&dir)
+                .log_err_and_ignore(&format!("Could not create domain dir {name}"));
         }
         Self {
             name: name.to_string(),

src/pages/mod.rs

@@ -14,7 +14,7 @@ use component::*;
 use serde_json::json;
 
 use crate::{
-    ai::{generate_embedding, DocEmbedding, EmbedStore},
+    ai::{generate_embedding, EmbedStore, SearchResult},
     archive::WebsiteArchive,
 };
 
@@ -176,7 +176,7 @@ pub async fn render_website(
     None
 }
 
-pub fn gen_search_element(x: &DocEmbedding) -> PreEscaped<String> {
+pub fn gen_search_element(x: &SearchResult) -> PreEscaped<String> {
     html! {
         div class="text-xl font-bold mt-4 p-4 flex items-center w-full max-w-4xl max-h-40 mx-auto bg-neutral-800 shadow-md rounded-lg overflow-hidden border border-neutral-900 hover:cursor-pointer"
         hx-get=(format!("/d/{}/{}", x.domain, x.path))
@@ -186,7 +186,7 @@ pub fn gen_search_element(x: &DocEmbedding) -> PreEscaped<String> {
                 a { (x.domain) };
                 (slash_seperator());
                 (gen_path_header(x.path.split('/').collect(), &x.domain, false));
-                p class="font-bold p-2 text-stone-400" { (format!("{:.2} %", x.similarity * 100.0)) };
+                p class="font-bold p-2 text-stone-400" { (format!("{:.2} % [{} matches]", x.similarity() * 100.0, x.chunks.len())) };
         };
     }
 }
@@ -220,13 +220,14 @@ pub async fn vector_search(
                     EmbedStore::search_vector(&input, limit as i64, offset as i64).await
                 })
             },
-            5,
+            50,
         )
         .pager(page as u64, vector)
         .await;
 
         // API Route
         if query.ends_with(".json") {
+            // TODO : Better search API
             return Some(respond_json(&json!(&results.page(page as u64))));
         }