From 785a46e5a3c3e52d78801f2ab1aecf3f8655062d Mon Sep 17 00:00:00 2001
From: Victor Giers <mail@victorgiers.com>
Date: Thu, 19 Mar 2026 22:33:41 +0100
Subject: [PATCH] Add ollama_embeddings.py for embedding functionality

---
 backend/rag/ollama_embeddings.py | 87 ++++++++++++++++++++++++++++++++
 1 file changed, 87 insertions(+)
 create mode 100644 backend/rag/ollama_embeddings.py

diff --git a/backend/rag/ollama_embeddings.py b/backend/rag/ollama_embeddings.py
new file mode 100644
index 0000000..fae9da5
--- /dev/null
+++ b/backend/rag/ollama_embeddings.py
@@ -0,0 +1,87 @@
+from __future__ import annotations
+
+from typing import Dict, Iterable, List, Optional, Tuple
+
+import requests
+
+
+DEFAULT_EMBED_CANDIDATES = (
+    "bge-m3:latest",
+    "nomic-embed-text:latest",
+    "dengcao/Qwen3-Embedding-0.6B:F16",
+)
+
+_MODEL_CACHE: Dict[Tuple[str, str], str] = {}
+
+
+def _cache_key(ollama_url: str, preferred_model: Optional[str]) -> Tuple[str, str]:
+    return (ollama_url.rstrip("/"), str(preferred_model or "").strip())
+
+
+def _candidate_models(preferred_model: Optional[str]) -> List[str]:
+    out: List[str] = []
+    seen = set()
+    for model in [preferred_model, *DEFAULT_EMBED_CANDIDATES]:
+        name = str(model or "").strip()
+        if not name or name in seen:
+            continue
+        out.append(name)
+        seen.add(name)
+    return out
+
+
+def _response_error_text(response: requests.Response) -> str:
+    try:
+        payload = response.json()
+    except Exception:
+        payload = None
+    if isinstance(payload, dict):
+        message = str(payload.get("error") or "").strip()
+        if message:
+            return message
+    return (response.text or f"HTTP {response.status_code}").strip()
+
+
+def request_embedding(ollama_url: str, model: str, text: str, *, timeout: int = 120) -> List[float]:
+    response = requests.post(
+        f"{ollama_url.rstrip('/')}/api/embeddings",
+        json={"model": model, "prompt": text},
+        timeout=timeout,
+    )
+    if not response.ok:
+        raise RuntimeError(f"{model}: {_response_error_text(response)}")
+
+    data = response.json()
+    vec = data.get("embedding") or (data.get("embeddings") or [None])[0]
+    if vec is None:
+        raise RuntimeError(f"{model}: Ollama returned no embedding vector")
+    return vec
+
+
+def resolve_embed_model(
+    ollama_url: str,
+    preferred_model: Optional[str],
+    *,
+    probe_text: str = "embedding probe",
+    timeout: int = 120,
+) -> Tuple[str, List[float]]:
+    key = _cache_key(ollama_url, preferred_model)
+    cached = _MODEL_CACHE.get(key)
+    if cached:
+        try:
+            return cached, request_embedding(ollama_url, cached, probe_text, timeout=timeout)
+        except Exception:
+            _MODEL_CACHE.pop(key, None)
+
+    errors: List[str] = []
+    for model in _candidate_models(preferred_model):
+        try:
+            vector = request_embedding(ollama_url, model, probe_text, timeout=timeout)
+            _MODEL_CACHE[key] = model
+            return model, vector
+        except Exception as exc:
+            errors.append(str(exc))
+
+    tried = ", ".join(_candidate_models(preferred_model)) or "(none)"
+    detail = "; ".join(errors) if errors else "no candidate models were available"
+    raise RuntimeError(f"No working Ollama embedding model found. Tried: {tried}. {detail}")