serve: add /api/embed, /api/create, /api/copy, /api/blobs; fix 415, 501

README.md

@@ -158,7 +158,7 @@ The server listens on `127.0.0.1:17434` by default, overridable via `LLMMAN_HOST
 
 | API | Endpoints |
 |-----|-----------|
-| Ollama | `/api/generate`, `/api/chat`, `/api/tags`, `/api/show`, `/api/pull`, `/api/ps`, `/api/delete` |
+| Ollama | `/api/generate`, `/api/chat`, `/api/embed`, `/api/embeddings`, `/api/tags`, `/api/show`, `/api/pull`, `/api/push`, `/api/copy`, `/api/create`, `/api/blobs/{digest}`, `/api/ps`, `/api/delete`, `/api/version` |
 | OpenAI | `/v1/chat/completions`, `/v1/completions`, `/v1/embeddings`, `/v1/models`, `/v1/responses`, `/v1/responses/input_tokens` |
 | Anthropic | `/v1/messages` |
 | llmman | `/llmman/providers`, `/llmman/providers/{id}` |
@@ -201,6 +201,15 @@ as `message.tool_calls`), `images` (vision, base64, same as Ollama's own
 wire format), and `format` (`"json"` or a JSON Schema object, for
 constrained structured output).
 
+`/api/embed` and `/api/embeddings` work with any embedding model (a GGUF
+with a pooling type, e.g. `embeddinggemma`, `nomic-embed-text`): `llama-server`
+is started with `--embeddings` for it, so `/v1/embeddings` works too.
+
+`/api/create` supports `from` (alias a model) and `files` (GGUFs uploaded via
+`/api/blobs/{digest}`, as `ollama create` does). Modelfile fields such as
+`system` or `quantize` are refused with a 400: the GGUF's own chat template
+applies.
+
 An idle, unused model is automatically unloaded after `keep_alive`
 (default 5 minutes, matching Ollama; set per-request, or daemon-wide via
 `LLMMAN_KEEP_ALIVE`), and `llmman ps`/`/api/ps` reports each model's

src/cmd/build.rs

@@ -22,10 +22,9 @@ pub struct BuildArgs {
 }
 
 pub fn run(args: &BuildArgs) -> anyhow::Result<()> {
-    // The build tag is a user-chosen label stored verbatim, not resolved, so
-    // this is its only validation gate. Check it before touching the store so
-    // a bad tag never creates the store tree.
-    crate::shortnames::validate_reference(&args.tag)?;
+    // Resolved the way run/rm/show look it up (see cp.rs), before
+    // touching the store so a bad tag never creates the store tree.
+    let tag = crate::shortnames::resolve_ollama_api(&args.tag)?;
 
     let store_root = crate::default_store()?;
     let store = OciStore::open(&store_root)?;
@@ -46,7 +45,7 @@ pub fn run(args: &BuildArgs) -> anyhow::Result<()> {
         .canonicalize()
         .with_context(|| format!("context dir: {}", args.context_dir.display()))?;
 
-    let desc = store.build(&context_dir, &args.tag, &labels)?;
+    let desc = store.build(&context_dir, &tag, &labels)?;
     println!("Built {} ({})", args.tag, desc.digest);
     Ok(())
 }

src/cmd/cp.rs

@@ -21,27 +21,32 @@ pub struct CpArgs {
 /// to match ollama's naming for anyone coming from there, not because the
 /// underlying operation differs at all.
 pub fn run(args: &CpArgs) -> anyhow::Result<()> {
-    // Validate/resolve both refs before touching the store, so a bad ref
-    // never creates the store tree. DESTINATION is stored verbatim, not
-    // resolved/canonicalized, so this is its only client-side validation
-    // gate. resolve_ollama_api for SOURCE, not resolve: it must match how
-    // the model is actually stored, same reasoning as tag.rs/rm.rs.
-    crate::shortnames::validate_reference(&args.destination)?;
+    // Both via resolve_ollama_api, before touching the store: SOURCE to
+    // match how it's stored, DESTINATION to be stored the way run/rm/show
+    // look it up (a bare `mine:v1` becomes `docker.io/ai/mine:v1`).
     let source = crate::shortnames::resolve_ollama_api(&args.source)?;
+    let destination = crate::shortnames::resolve_ollama_api(&args.destination)?;
 
     let store_root = crate::default_store()?;
     let store = OciStore::open(&store_root)?;
+    copy(&store, &source, &destination)?;
+    println!("copied '{}' to '{}'", args.source, args.destination);
+    Ok(())
+}
 
-    let desc = store.find(&source)?;
+/// Points `destination` at the content `source` names (both already
+/// resolved) and returns its digest. Shared by `llmman cp` and `/api/copy`.
+pub fn copy(store: &OciStore, source: &str, destination: &str) -> anyhow::Result<String> {
+    let desc = store.find(source)?;
     // A destination spelled as a digest names content, and the store
     // answers such a reference by content (see `OciStore::find`): a
     // pointer there holding some other digest would be passed over by a
     // lookup for the digest it is named for and answer only for the one
     // it holds, so it is refused.
-    check_digest_destination(&args.destination, &desc.digest)?;
-    store.tag(desc, &args.destination)?;
-    println!("copied '{}' to '{}'", args.source, args.destination);
-    Ok(())
+    check_digest_destination(destination, &desc.digest)?;
+    let digest = desc.digest.clone();
+    store.tag(desc, destination)?;
+    Ok(digest)
 }
 
 /// Refuses a `<name>@<digest>` destination whose digest is not `actual`,