[OMEGA-400] ASI Cloud embeddings support

Autotests/test_openai_runtime_embeddings.py

@@ -9,12 +9,13 @@
 MEMORY_METTA_PATH = REPO_ROOT / "src" / "memory.metta"
 
 
-def load_rag_module(monkeypatch):
+def load_rag_module(monkeypatch, config=None, expected_model="text-embedding-3-large"):
     created_clients = []
+    settings = {"GATEWAY_URL": "http://gateway:8080", **(config or {})}
 
     class FakeEmbeddings:
         def create(self, *, model, input):
-            assert model == "text-embedding-3-large"
+            assert model == expected_model
             assert input == ["runtime probe"]
             return types.SimpleNamespace(
                 data=[types.SimpleNamespace(embedding=[0.1, 0.2, 0.3])]
@@ -32,7 +33,7 @@ def __init__(self, *, base_url=None, api_key=None):
     chromadb_module = types.ModuleType("chromadb")
     config_module = types.ModuleType("config")
     config_module.config_get_by_key = (
-        lambda key, default=None: "http://gateway:8080" if key == "GATEWAY_URL" else default
+        lambda key, default=None: settings.get(key, default)
     )
     llm_module = types.ModuleType("lib_llm_ext")
     llm_module.initLocalEmbedding = lambda: None
@@ -59,6 +60,18 @@ def test_runtime_openai_embedding_uses_proxy_and_returns_single_vector(monkeypat
     assert clients[0].api_key == "unused"
 
 
+def test_runtime_embedding_uses_the_configured_provider_and_model(monkeypatch):
+    rag, clients = load_rag_module(
+        monkeypatch,
+        config={"embeddingprovider": "ASICloud",
+                "embedding_model": "WhereIsAI/UAE-Large-V1"},
+        expected_model="WhereIsAI/UAE-Large-V1",
+    )
+
+    assert rag.openai_embed("runtime probe") == [0.1, 0.2, 0.3]
+    assert clients[0].base_url == "http://gateway:8080/asicloud/"
+
+
 def test_memory_metta_routes_openai_embeddings_to_rag_wrapper():
     memory_metta = MEMORY_METTA_PATH.read_text(encoding="utf-8")
 

config/config.yaml

@@ -34,8 +34,10 @@ maxRecallItems: 20
 maxEpisodeRecallLines: 20
 # Tail of `memory/history.metta` included in the prompt (chars)
 maxHistory: 30000
-# `Local` (Python-side model) or `OpenAI` (requires `OPENAI_API_KEY`)
+# `Local` (Python-side model) or the id of a provider serving OpenAI-compatible embeddings: `OpenAI`, `ASICloud`
 embeddingprovider: Local
+# Model asked of a non-`Local` embeddingprovider
+embedding_model: "text-embedding-3-large"
 
 # Policy
 

docs/reference-configuration.md

@@ -31,7 +31,8 @@ This reads a command-line override via `argk` (`name=value` on the MeTTa command
 | `maxRecallItems` | 20 | Items returned by `query`. |
 | `maxEpisodeRecallLines` | 20 | Lines returned by `episodes`. |
 | `maxHistory` | 30000 (chars) | Tail of `memory/history.metta` included in the prompt. |
-| `embeddingprovider` | `Local` | `Local` (Python-side model) or `OpenAI`. |
+| `embeddingprovider` | `Local` | `Local` (Python-side model), or the id of a provider that serves an OpenAI-compatible `/embeddings` endpoint — `OpenAI` and `ASICloud` are known to. The gateway supplies that provider's key. |
+| `embedding_model` | `text-embedding-3-large` | Model asked of a non-`Local` `embeddingprovider`. |
 
 ## Channels (`src/channels.metta`, `initChannels`)
 

docs/reference-internals-extension-points.md

@@ -53,10 +53,14 @@ In `src/memory.metta`, the `embed` function dispatches on `embeddingprovider`:
 (= (embed $str)
    (if (== (embeddingprovider) Local)
        (py-call (lib_llm_ext.useLocalEmbedding (string-safe $str)))
-       (useGPTEmbedding (string-safe $str))))
+       (py-call (rag.openai_embed (string-safe $str)))))
 ```
 
-To add a new backend, add a branch and implement the Python function.
+Any value other than `Local` is a provider id: the remote branch posts
+`embedding_model` to `<GATEWAY_URL>/<embeddingprovider lowercased>/`, the
+location that already injects that provider's key. Switching vendor is
+configuration, not code — provided the vendor serves embeddings at all. It
+changes the vector space, so reset the ChromaDB store when you do.
 
 ## Change the reasoning library
 

docs/reference-skills-memory.md

@@ -30,7 +30,7 @@ The result of the ChromaDB write (internally). The agent treats a successful cal
 
 ### Notes / Limits
 - Text is passed through `string-safe` before embedding, which escapes newlines, quotes, and apostrophes.
-- Embedding provider is selected by `embeddingprovider` (`Local` or `OpenAI`).
+- Embedding provider is selected by `embeddingprovider`, the model by `embedding_model`.
 - Nothing deduplicates automatically — repeated `remember` calls store multiple items.
 
 ---

scripts/omegaclaw

@@ -497,6 +497,10 @@ help() {
     echo -e "\t-l <logging config> set Python logging config file"
     echo -e "\t--version, -v show the OmegaClaw version"
     echo -e "\t--help, -h show this help"
+    echo
+    echo -e "Environment:"
+    echo -e "\tEMBEDDING_PROVIDER override the embedding backend (default depends on provider)"
+    echo -e "\tEMBEDDING_MODEL set the model asked of a non-Local EMBEDDING_PROVIDER"
 }
 
 options() {
@@ -586,6 +590,10 @@ options() {
 }
 
 start() {
+    # Each provider above picks a default embedding backend. EMBEDDING_PROVIDER
+    # overrides it, so a remote provider can serve embeddings too.
+    embeddingprovider="${EMBEDDING_PROVIDER:-${embeddingprovider}}"
+
     docker rm -f omegaclaw 2>/dev/null || true
     docker pull "${image}" 2>/dev/null || true
 
@@ -656,6 +664,10 @@ start() {
       docker_cmd+=("openaiapi_url=${openaiapi_url}")
     fi
 
+    if [ -n "${EMBEDDING_MODEL:-}" ]; then
+      docker_cmd+=("embedding_model=${EMBEDDING_MODEL}")
+    fi
+
     if [ -n "${openclaw_url:-}" ]; then
       docker_cmd+=("openclaw_url=${openclaw_url}")
       docker_cmd+=("openClawEnabled=enabled")

src/loop.metta

@@ -26,9 +26,9 @@
 
 (= (initKnowledge)
    (progn (log INFO "loop" "Initializing knowledge base")
-          (if (== (embeddingprovider) OpenAI)
-               (log INFO "loop" (py-call (rag.init_knowledge "OpenAI")))
-               (log INFO "loop" (py-call (rag.init_knowledge "Local"))))))
+          (if (== (embeddingprovider) Local)
+               (log INFO "loop" (py-call (rag.init_knowledge "Local")))
+               (log INFO "loop" (py-call (rag.init_knowledge "OpenAI"))))))
 
 (= (getContext)
    (string-safe (py-str ("PROMPT: " (getPrompt (provider)) " SKILLS: " (getSkills)

src/rag.py

@@ -139,14 +139,16 @@ def _chunk_markdown(text, filename):
 # --- Embedding -----------------------------------------------------------
 
 def openai_embed_batch(texts):
-    """Embed a list of texts via OpenAI. Returns list of float vectors."""
+    """Embed a list of texts via an OpenAI-compatible API. Returns list of float vectors."""
+    model = config_get_by_key("embedding_model", EMBEDDING_MODEL)
     proxy_url = config_get_by_key("GATEWAY_URL")
     if proxy_url:
-        client = openai.OpenAI(base_url=f"{proxy_url.rstrip('/')}/openai/", api_key="unused")
+        prefix = str(config_get_by_key("embeddingprovider", "OpenAI")).lower()
+        client = openai.OpenAI(base_url=f"{proxy_url.rstrip('/')}/{prefix}/", api_key="unused")
     else:
         client = openai.OpenAI()
     try:
-        resp = client.embeddings.create(model=EMBEDDING_MODEL, input=texts)
+        resp = client.embeddings.create(model=model, input=texts)
     except Exception as e:
         raise RuntimeError(f"Embedding request failed: {e}") from e
     return [item.embedding for item in resp.data]