diff --git a/app/ai_analyzer.py b/app/ai_analyzer.py
index 867300ef..5bbe835c 100644
--- a/app/ai_analyzer.py
+++ b/app/ai_analyzer.py
@@ -87,6 +87,22 @@ def rewrite_ai_url_for_runtime(url: str) -> str:
     return rewritten
 
 
+def openai_compat_base_url(url: str, endpoint: str = "/chat/completions") -> str:
+    """Return the OpenAI-compatible *base* URL (scheme+host+optional /v1).
+
+    Users often paste the full chat URL (…/v1/chat/completions) into the
+    Settings base-URL field. Built-in providers then append ``endpoint``
+    again and Ollama returns 404. Strip a trailing copy of ``endpoint``.
+    """
+    if not url:
+        return url
+    base = url.rstrip("/")
+    ep = "/" + (endpoint or "").lstrip("/")
+    if ep not in ("", "/") and base.lower().endswith(ep.lower()):
+        base = base[: -len(ep)].rstrip("/")
+    return base
+
+
 def dummy_key_for_local(api_key: Optional[str]) -> str:
     """Ollama accepts any Bearer token; the proxy still sends Authorization."""
     key = (api_key or "").strip()
@@ -2541,7 +2557,8 @@ def call_ai_provider(provider: str,
         base = api_url.rstrip('/')
         full_url = f"{base}/models/{model}:generateContent"
     else:
-        full_url = api_url.rstrip('/') + '/' + endpoint.lstrip('/')
+        base = openai_compat_base_url(api_url, endpoint)
+        full_url = base.rstrip('/') + '/' + endpoint.lstrip('/')
     
     return _execute_ai_request(full_url, headers, ai_request_payload)
 
diff --git a/app/app_base.py b/app/app_base.py
index 46536dac..8a007b04 100644
--- a/app/app_base.py
+++ b/app/app_base.py
@@ -1343,7 +1343,8 @@ def ai_api_call():
             else:
                 full_url = f"{api_url.rstrip('/')}/models/{model}:generateContent"
         else:
-            full_url = api_url.rstrip('/') + '/' + endpoint.lstrip('/')
+            base = ai_analyzer.openai_compat_base_url(api_url, endpoint)
+            full_url = base.rstrip('/') + '/' + endpoint.lstrip('/')
         logger.debug("Full URL: %s", full_url)
         
         # Prepare headers
diff --git a/app/assets/js/settings.js b/app/assets/js/settings.js
index 4eae3417..109ef27e 100644
--- a/app/assets/js/settings.js
+++ b/app/assets/js/settings.js
@@ -121,7 +121,7 @@ const AI_PROVIDERS = [
         defaultUrl: 'http://localhost:11434/v1',
         apiKeyOptional: true,
         allowCustomModel: true,
-        helpHtml: 'Talks to Ollama, LM Studio, vLLM, or llama.cpp on <em>your laptop</em> (not inside Docker). From Docker Desktop, <code>localhost</code> is rewritten to <code>host.docker.internal</code>. Ollama default port is <code>11434</code>; LM Studio is usually <code>1234</code>.',
+        helpHtml: 'Use the <strong>base URL only</strong> — <code>http://localhost:11434/v1</code>. Do <em>not</em> add <code>/chat/completions</code> (the app appends it). From Docker Desktop, <code>localhost</code> is rewritten to <code>host.docker.internal</code>. Ollama port <code>11434</code>; LM Studio is usually <code>1234</code>.',
         models: [
             { id: 'llama3.2', name: 'llama3.2' },
             { id: 'llama3.1', name: 'llama3.1' },
diff --git a/tests/python/test_local_openai.py b/tests/python/test_local_openai.py
index c5e56b42..02f524ae 100644
--- a/tests/python/test_local_openai.py
+++ b/tests/python/test_local_openai.py
@@ -2,6 +2,7 @@
 from ai_analyzer import (
     dummy_key_for_local,
     is_local_openai_compat,
+    openai_compat_base_url,
     rewrite_ai_url_for_runtime,
     running_in_docker,
 )
@@ -40,6 +41,18 @@ def test_rewrite_localhost_inside_docker(monkeypatch):
     )
 
 
+def test_openai_compat_base_url_strips_chat_completions():
+    assert openai_compat_base_url("http://localhost:11434/v1") == "http://localhost:11434/v1"
+    assert (
+        openai_compat_base_url("http://host.docker.internal:11434/v1/chat/completions")
+        == "http://host.docker.internal:11434/v1"
+    )
+    assert (
+        openai_compat_base_url("http://localhost:11434/v1/chat/completions/")
+        == "http://localhost:11434/v1"
+    )
+
+
 def test_running_in_docker_env(monkeypatch):
     monkeypatch.setenv("CBQA_IN_DOCKER", "true")
     assert running_in_docker() is True