daemon: embeds through the endpoint pool, GPUs together and the CPU last

query-embed-on-cpu, follow-up daemon-embedder-gpu-pool: huginn-daemon takes
repeated --ollama-url (GPU hosts) and --reserve-ollama-url (CPU hosts) and
builds the same EmbedderPool as huginn-sources, through one ollama::pool
constructor. A lone --ollama-url is a pool of one.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

README.md

@@ -26,7 +26,9 @@ Upstream: `https://github.com/GameCult/Huginn.git`
   restart compares the mind with the collection and writes what is missing. An
   unreachable Qdrant or Ollama shows in `whoami` as the index's status and is
   retried; admission never waits on it. The daemon requires `--qdrant-url`,
-  `--ollama-url` and `--embedding-model`. A `query` carrying `semantic` is
+  `--ollama-url` and `--embedding-model`; `--ollama-url` repeats for each GPU
+  host and `--reserve-ollama-url` for each CPU host, and the GPU hosts embed
+  together while the CPU hosts serve only when none can. A `query` carrying `semantic` is
   answered through the same selection as any other: the index supplies
   candidate ids and scores, and the mind keeps only the documents it holds and
   the selection admits (ask for `in_force` to exclude resolved ones), ordered by

crates/huginn-daemon/src/bin/huginn-sources.rs

@@ -7,8 +7,7 @@ use std::path::PathBuf;
 use std::process::ExitCode;
 
 use anyhow::{Context, Result, ensure};
-use huginn_daemon::index::ollama::OllamaEmbedder;
-use huginn_daemon::index::pool::EmbedderPool;
+use huginn_daemon::index::ollama::pool;
 use huginn_daemon::index::qdrant::QdrantIndex;
 use huginn_daemon::sources::forge::HttpForge;
 use huginn_daemon::sources::refresh;
@@ -77,8 +76,7 @@ fn main() -> ExitCode {
             return ExitCode::from(2);
         }
     };
-    let endpoint = |url: &String| (url.clone(), OllamaEmbedder::new(url, &options.embedding_model));
-    let mut embedder = EmbedderPool::new(options.ollama_urls.iter().map(endpoint).collect(), options.reserve_urls.iter().map(endpoint).collect());
+    let mut embedder = pool(&options.ollama_urls, &options.reserve_urls, &options.embedding_model);
     let mut index = QdrantIndex::new(&options.qdrant_url);
     let mut forge = HttpForge::new(&options.forge_url);
     let mut journal = |report: &huginn_daemon::sources::Report| println!("{report}");

crates/huginn-daemon/src/index/ollama.rs

@@ -73,3 +73,11 @@ impl Embedder for OllamaEmbedder {
             .collect()
     }
 }
+
+/// One pool over Ollama endpoints serving `model`: the `primaries` (GPU hosts)
+/// share the work and the `reserve` (CPU hosts) serve only when none of them
+/// can. The one embedder client of every Huginn process that embeds.
+pub fn pool(primaries: &[String], reserve: &[String], model: &str) -> super::pool::EmbedderPool<OllamaEmbedder> {
+    let endpoint = |url: &String| (url.clone(), OllamaEmbedder::new(url, model));
+    super::pool::EmbedderPool::new(primaries.iter().map(endpoint).collect(), reserve.iter().map(endpoint).collect())
+}

crates/huginn-daemon/src/main.rs

@@ -7,13 +7,13 @@ use std::sync::atomic::AtomicBool;
 
 use anyhow::{Context, Result};
 use huginn_daemon::index::Backoff;
-use huginn_daemon::index::ollama::OllamaEmbedder;
+use huginn_daemon::index::ollama::pool;
 use huginn_daemon::index::qdrant::QdrantIndex;
 use huginn_daemon::serve::{ServeOptions, parse_options, run, startup};
 
 fn main() -> Result<()> {
     let options = parse_options(std::env::args().skip(1))?;
-    let embedder = OllamaEmbedder::new(&options.ollama_url, &options.embedding_model);
+    let embedder = pool(&options.ollama_urls, &options.reserve_ollama_urls, &options.embedding_model);
     let index = QdrantIndex::new(&options.qdrant_url);
     let (mut daemon, mut hub, registry) = match startup(&options, embedder, index, Backoff::default()) {
         Ok(started) => started,

crates/huginn-daemon/src/serve.rs

@@ -101,7 +101,7 @@ impl Default for ServeOptions {
     }
 }
 
-/// The six things the operator supplies. Nothing here reads the environment:
+/// The things the operator supplies. Nothing here reads the environment:
 /// how Idunn supplies `--bind` is Cut 14's. The index has no off switch: an
 /// unreachable Qdrant or Ollama is a degraded projection, reported by
 /// `whoami`, not a configuration.
@@ -111,23 +111,34 @@ pub struct Options {
     pub instance: Slug,
     pub bind: SocketAddr,
     pub qdrant_url: String,
-    pub ollama_url: String,
+    /// The GPU hosts: all take embedding at once. One is the plain case.
+    pub ollama_urls: Vec<String>,
+    /// The CPU host: takes embedding only while no GPU host can.
+    pub reserve_ollama_urls: Vec<String>,
     pub embedding_model: String,
 }
 
 /// Exactly `--state-root`, `--instance`, `--bind`, `--qdrant-url`,
-/// `--ollama-url` and `--embedding-model`, each required, each once. The two
-/// URLs are plain `http://`: this build carries no TLS.
+/// `--ollama-url` and `--embedding-model`, each required, each once, except
+/// `--ollama-url`, which repeats for each GPU host, and the optional
+/// `--reserve-ollama-url`, which repeats for each CPU host. The URLs are plain
+/// `http://`: this build carries no TLS.
 pub fn parse_options(args: impl Iterator<Item = String>) -> Result<Options> {
     let mut args = args;
     let mut values: BTreeMap<String, String> = BTreeMap::new();
+    let (mut ollama_urls, mut reserve_ollama_urls) = (Vec::new(), Vec::new());
     while let Some(name) = args.next() {
         let name = name.strip_prefix("--").with_context(|| format!("expected --option, got {name:?}"))?;
         ensure!(
-            matches!(name, "state-root" | "instance" | "bind" | "qdrant-url" | "ollama-url" | "embedding-model"),
+            matches!(name, "state-root" | "instance" | "bind" | "qdrant-url" | "ollama-url" | "reserve-ollama-url" | "embedding-model"),
             "unsupported Huginn option --{name}"
         );
         let value = args.next().with_context(|| format!("missing value for --{name}"))?;
+        if name == "ollama-url" || name == "reserve-ollama-url" {
+            ensure!(value.starts_with("http://"), "--{name} must be a plain http:// address, got {value:?}");
+            if name == "ollama-url" { ollama_urls.push(value) } else { reserve_ollama_urls.push(value) }
+            continue;
+        }
         ensure!(values.insert(name.to_owned(), value).is_none(), "duplicate Huginn option --{name}");
     }
     let take = |name: &str| -> Result<String> {
@@ -148,7 +159,11 @@ pub fn parse_options(args: impl Iterator<Item = String>) -> Result<Options> {
         instance: Slug(take("instance")?),
         bind: bind.parse().with_context(|| format!("--bind must be an ip:port, got {bind:?}"))?,
         qdrant_url: url("qdrant-url")?,
-        ollama_url: url("ollama-url")?,
+        ollama_urls: {
+            ensure!(!ollama_urls.is_empty(), "--ollama-url is required; repeat it for each GPU host");
+            ollama_urls
+        },
+        reserve_ollama_urls,
         embedding_model,
     })
 }
@@ -1347,7 +1362,8 @@ mod tests {
             instance: slug(INSTANCE),
             bind: port,
             qdrant_url: "http://127.0.0.1:1".into(),
-            ollama_url: "http://127.0.0.1:1".into(),
+            ollama_urls: vec!["http://127.0.0.1:1".into()],
+            reserve_ollama_urls: Vec::new(),
             embedding_model: "a-model".into(),
         };
         let started = |options: &Options| {
@@ -1369,7 +1385,7 @@ mod tests {
     }
 
     #[test]
-    fn options_are_exactly_the_six_the_operator_supplies() {
+    fn options_are_the_ones_the_operator_supplies() {
         let root = if cfg!(windows) { r"C:\state" } else { "/state" };
         let args = |values: &[&str]| values.iter().map(|value| value.to_string()).collect::<Vec<_>>().into_iter();
         let good = [
@@ -1383,7 +1399,8 @@ mod tests {
                 instance: slug("yggdrasil"),
                 bind: "127.0.0.1:17872".parse().unwrap(),
                 qdrant_url: "http://127.0.0.1:6333".into(),
-                ollama_url: "http://10.77.0.4:11434".into(),
+                ollama_urls: vec!["http://10.77.0.4:11434".into()],
+                reserve_ollama_urls: Vec::new(),
                 embedding_model: "qwen3-embedding:0.6b".into(),
             }
         );
@@ -1411,5 +1428,11 @@ mod tests {
         ] {
             assert!(parse_options(args(&bad)).is_err(), "{bad:?}");
         }
+        // The GPU hosts and the CPU host repeat; a lone --ollama-url is the plain case.
+        let pooled = [good.to_vec(), vec!["--ollama-url", "http://10.77.0.2:11434", "--reserve-ollama-url", "http://127.0.0.1:11434"]].concat();
+        let parsed = parse_options(args(&pooled)).unwrap();
+        assert_eq!(parsed.ollama_urls, ["http://10.77.0.4:11434", "http://10.77.0.2:11434"]);
+        assert_eq!(parsed.reserve_ollama_urls, ["http://127.0.0.1:11434"]);
+        assert!(parse_options(args(&[good.to_vec(), vec!["--reserve-ollama-url", "127.0.0.1:11434"]].concat())).is_err());
     }
 }