litellm: route local_qwen2.5-14b via openai/, not ollama_chat/

paperless-ai kept returning "Invalid JSON response from API" after the
real cause of the large failures was fixed (TOKEN_LIMIT=128000 against a
16386 window; requests fell from 44k to 2.1k tokens and truncation
stopped). Something in the hop was still mangling the reply.

Isolated by replaying paperless-ai's exact system prompt 3x per path,
identical model, temperature and context:
  Anvil /v1/chat/completions (OpenAI-compat) -> 3/3 VALID
  Anvil /api/chat            (native)        -> 3/3 VALID
  via LiteLLM ollama_chat/                   -> invalid every time

Both ollama endpoints are clean, so the fault is the ollama_chat provider
transformation. openai/ is a near-passthrough to the endpoint already
proven good.

Deliberately NOT enabling LiteLLM verbose logging to confirm this from the
inside: those request bodies are tax and medical documents, and writing
them to Loki would undo the privacy property that put paperless-ai on a
local model in the first place.

Scoped to this one route to keep the blast radius small; the other local_*
entries are untouched.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>

vars/main.yml

@@ -94,15 +94,28 @@ litellm_models:
     backend: ollama_chat/llama3.1:8b
     api_base: "{{ anvil_ollama_base }}"
     api_key: ollama-no-auth
-  # The structured-output tier, added 2026-09-12. llama3.1:8b answers fast but
+  # The structured-output tier for paperless-ai, added 2026-09-12. ~9 GB, so it
-  # will not hold strict JSON: paperless-ai got "Expected ',' or '}' after
+  # co-exists with anything else here.
-  # property value in JSON at position 263" on a 1,921-token reply. The Kestra
+  #
-  # flows escape this by setting response_format={"type":"json_object"};
+  # ⚠️ NOTE THE PROVIDER: `openai/` with an explicit /v1, NOT `ollama_chat/`
-  # paperless-ai does not, so it needs a model that keeps the shape unprompted.
+  # like every other local_* route above. This is deliberate and measured.
-  # 14B is the smallest that does, and at ~9 GB it co-exists with anything here.
+  #
+  # paperless-ai kept failing with "Invalid JSON response from API" even after
+  # the real cause of the big failures (TOKEN_LIMIT=128000 against a 16386
+  # window) was fixed and requests dropped from 44k to 2.1k tokens. Isolation
+  # on 2026-09-12, replaying paperless-ai's exact system prompt 3x per path:
+  #   Anvil /v1/chat/completions  (OpenAI-compat) → 3/3 VALID JSON
+  #   Anvil /api/chat             (native)        → 3/3 VALID JSON
+  #   via LiteLLM ollama_chat/                    → invalid every time
+  # Model, prompt, temperature and context were identical across all three, so
+  # the fault is the provider's transformation, not the model.
+  #
+  # `openai/` is a near-passthrough to the endpoint that was proven to work.
+  # ⛔ Do not "tidy" this back to ollama_chat/ for consistency with its
+  # neighbours. The inconsistency is the fix.
   - name: local_qwen2.5-14b
-    backend: ollama_chat/qwen2.5:14b
+    backend: openai/qwen2.5:14b
-    api_base: "{{ anvil_ollama_base }}"
+    api_base: "{{ anvil_ollama_base }}/v1"
     api_key: ollama-no-auth
   # Shadow alias SOLELY for paperless-ai's hardcoded gpt-4o-mini validation
   # probe (see the "Narrow exception" note above). Routed to free local