Bug Description

Bug Description When streaming responses from Ollama models that support extended thinking (e.g., DeepSeek-R1, QwQ), chunks where content=None but thinking is present are silently dropped. This means all thinking content is lost during streaming, even though think=True is properly passed to the Ollama API.

The non-streaming chat() and achat() methods handle thinking blocks correctly — only the streaming paths are affected.

Root Cause In llama_index/llms/ollama/base.py, both stream_chat and astream_chat have this guard:

if r["message"]["content"] is None: continue During the thinking phase, Ollama sends chunks where content is None and thinking contains the reasoning text. This guard skips those chunks entirely before the ThinkingBlock accumulation logic can process them.

Suggested Fix Change the guard in both stream_chat (~line 474) and astream_chat (~line 558) to:

if r["message"]["content"] is None and not r["message"].get("thinking"): continue This allows thinking-only chunks through while still skipping truly empty chunks.

Environment llama-index-llms-ollama: 0.10.1 llama-index-core: 0.14.19 Ollama model: DeepSeek-R1 (any model with think=True support)

Version

0.14.19

Steps to Reproduce

Steps to Reproduce Install packages

pip install llama-index-llms-ollama==0.10.1 llama-index-core==0.14.19
Pull an Ollama model that supports thinking (e.g., DeepSeek-R1)

ollama pull deepseek-r1 Run this script:

from llama_index.llms.ollama import Ollama
from llama_index.core.llms import ChatMessage

llm = Ollama(model="deepseek-r1", request_timeout=120, thinking=True)
messages = [ChatMessage(role="user", content="What is 25 * 25?")]

# --- Non-streaming (works correctly) ---
response = llm.chat(messages)
print("Thinking:", response.additional_kwargs.get("thinking"))  # ✅ Has thinking content
print("Answer:", response.message.content)

# --- Streaming (thinking is lost) ---
thinking_chunks = []
text_chunks = []
for chunk in llm.stream_chat(messages):
    thinking = (chunk.additional_kwargs or {}).get("thinking_delta", "")
    if thinking:
        thinking_chunks.append(thinking)
    if chunk.delta:
        text_chunks.append(chunk.delta)

print("Thinking chunks received:", len(thinking_chunks))  # ❌ Always 0
print("Text chunks received:", len(text_chunks))

Expected Behavior thinking_chunks should contain the model's reasoning tokens, just like the non-streaming chat() returns them in additional_kwargs["thinking"].

Actual Behavior thinking_chunks is always empty. All chunks where content=None (which is every chunk during the thinking phase) are dropped by the continue guard before thinking_delta is ever set.

Relevant Logs/Tracbacks