Overview
Thinking mode enables orchid01 to reason through complex problems before responding. The model’s reasoning process is returned separately in reasoning_content alongside the final answer in content.
Use thinking mode for tasks that require multi-step analysis — complex covenant structures, cross-document comparison, scenario modelling. For straightforward extraction or Q&A, standard mode is faster and sufficient.
Enabling thinking
Pass thinking: true in the orchid config object:
response = client.chat.completions.create(
model="orchid01",
messages=[{
"role": "user",
"content": "Analyse the covenant package across these three credit agreements and identify any cross-default provisions..."
}],
max_tokens=16384,
extra_body={"orchid": {"thinking": True}},
)
reasoning = response.choices[0].message.model_extra.get("reasoning_content", "")
answer = response.choices[0].message.content
print("Reasoning:", reasoning)
print("Answer:", answer)
Streaming with thinking
Both reasoning and answer stream separately. Each chunk includes a type field:
stream = client.chat.completions.create(
model="orchid01",
messages=[...],
stream=True,
extra_body={"orchid": {"thinking": True}},
)
for chunk in stream:
delta = chunk.choices[0].delta
# reasoning chunks come first
reasoning_piece = getattr(delta, "reasoning_content", None)
if reasoning_piece:
print(reasoning_piece, end="", flush=True)
# then the final answer
content_piece = delta.content or ""
if content_piece:
print(content_piece, end="", flush=True)
Token requirements
Thinking mode requires max_tokens ≥ 16,000. Reasoning tokens count toward the limit — the model may be cut off mid-reasoning without enough headroom.
| Scenario | Recommended max_tokens |
|---|
| Short analysis | 16384 |
| Multi-document analysis | 32768 |
| Complex agentic workflows | 32768 |
max_tokens below 16,000 with thinking enabled, Orchid raises it automatically and sets max_tokens_adjusted: true in the response.
Temperature
Temperature is fixed at 1.0 in thinking mode — you cannot override it. Orchid sets this automatically so you don’t need to pass it.
When to use thinking mode
| Use thinking | Use standard |
|---|
| Complex covenant analysis across multiple docs | Single document extraction |
| Cross-default and cross-acceleration identification | Revenue and metric extraction |
| Multi-step scenario modelling | Filing summarisation |
| Ambiguous regulatory interpretation | Structured data conversion |
