Skip to content

Commit f132b34

Browse files
Algorithm5838github-actions[bot]
Algorithm5838
authored and
github-actions[bot]
committed
fix: preserve reasoning_details for multi-turn continuity
1 parent 2d54977 commit f132b34

2 files changed

Lines changed: 144 additions & 22 deletions

File tree

‎backend/open_webui/utils/middleware.py‎

Lines changed: 118 additions & 14 deletions
Original file line numberDiff line numberDiff line change
@@ -204,6 +204,35 @@ def split_json_objects(raw: str) -> list[str]:
204204
return expanded
205205

206206

207+
def _merge_reasoning_details(details: list, chunk: object) -> None:
208+
"""Merge reasoning_details delta items into a list in-place.
209+
210+
Items sharing an integer index are merged: text and summary deltas
211+
concatenate when string-valued, other fields overwrite with the latest
212+
delta. Items without a usable index are appended in arrival order.
213+
"""
214+
items = chunk if isinstance(chunk, list) else ([chunk] if isinstance(chunk, dict) else [])
215+
for item in items:
216+
if not isinstance(item, dict):
217+
continue
218+
idx = item.get('index')
219+
existing = None
220+
if isinstance(idx, int) and idx >= 0:
221+
existing = next((e for e in details if e.get('index') == idx), None)
222+
if existing is None:
223+
new_entry = dict(item)
224+
new_entry.setdefault('type', 'reasoning.text')
225+
details.append(new_entry)
226+
continue
227+
for k, v in item.items():
228+
if k in ('text', 'summary'):
229+
if isinstance(v, str):
230+
base = existing.get(k)
231+
existing[k] = (base + v) if isinstance(base, str) else v
232+
else:
233+
existing[k] = v
234+
235+
207236
def get_citation_source_from_tool_result(
208237
tool_name: str, tool_params: dict, tool_result: str, tool_id: str = ''
209238
) -> list[dict]:
@@ -508,6 +537,8 @@ def serialize_output(output: list) -> str:
508537
pass
509538

510539
reasoning_content = ''.join(reasoning_parts).strip()
540+
if not reasoning_content:
541+
continue # no displayable text; item stays in output for round-trip
511542

512543
duration = item.get('duration')
513544
status = item.get('status', 'in_progress')
@@ -3496,15 +3527,43 @@ async def non_streaming_chat_response_handler(response, ctx):
34963527
# otherwise generate from response content
34973528
response_output = response_data.get('output')
34983529
if not response_output:
3499-
response_output = [
3530+
message_obj = choices[0].get('message', {})
3531+
reasoning_text = (
3532+
message_obj.get('reasoning_content')
3533+
or message_obj.get('reasoning')
3534+
)
3535+
reasoning_details = message_obj.get('reasoning_details')
3536+
3537+
response_output = []
3538+
3539+
if reasoning_text or reasoning_details:
3540+
r_item = {
3541+
'type': 'reasoning',
3542+
'id': output_id('r'),
3543+
'status': 'completed',
3544+
'start_tag': '<think>',
3545+
'end_tag': '</think>',
3546+
'attributes': {'type': 'reasoning_content'},
3547+
'content': [{'type': 'output_text', 'text': reasoning_text}] if reasoning_text else [],
3548+
'summary': None,
3549+
}
3550+
if reasoning_details:
3551+
r_item['reasoning_details'] = (
3552+
reasoning_details
3553+
if isinstance(reasoning_details, list)
3554+
else [reasoning_details]
3555+
)
3556+
response_output.append(r_item)
3557+
3558+
response_output.append(
35003559
{
35013560
'type': 'message',
35023561
'id': output_id('msg'),
35033562
'status': 'completed',
35043563
'role': 'assistant',
35053564
'content': [{'type': 'output_text', 'text': content}],
35063565
}
3507-
]
3566+
)
35083567

35093568
await event_emitter(
35103569
{
@@ -3854,6 +3913,8 @@ def set_last_text(out, text):
38543913

38553914
# Initialize output: use existing from message if continuing, else create new
38563915
existing_output = message.get('output') if message else None
3916+
_pending_reasoning_details = []
3917+
38573918
if existing_output:
38583919
output = existing_output
38593920
else:
@@ -4237,9 +4298,14 @@ async def flush_pending_delta_data(threshold: int = 0):
42374298
or delta.get('reasoning')
42384299
or delta.get('thinking')
42394300
)
4301+
reasoning_details_chunk = delta.get('reasoning_details')
4302+
4303+
# Only create a reasoning item for visible reasoning text.
4304+
# Details-only deltas (e.g. Gemini encrypted blobs) are
4305+
# buffered to avoid splitting the assistant message mid-stream.
42404306
if reasoning_content:
42414307
if not output or output[-1].get('type') != 'reasoning':
4242-
reasoning_item = {
4308+
output.append({
42434309
'type': 'reasoning',
42444310
'id': output_id('r'),
42454311
'status': 'in_progress',
@@ -4249,23 +4315,37 @@ async def flush_pending_delta_data(threshold: int = 0):
42494315
'content': [],
42504316
'summary': None,
42514317
'started_at': time.time(),
4252-
}
4253-
output.append(reasoning_item)
4254-
else:
4255-
reasoning_item = output[-1]
4318+
})
42564319

4257-
# Append to reasoning content
4320+
if reasoning_content:
4321+
reasoning_item = output[-1]
42584322
parts = reasoning_item.get('content', [])
42594323
if parts and parts[-1].get('type') == 'output_text':
42604324
parts[-1]['text'] += reasoning_content
42614325
else:
4262-
reasoning_item['content'] = [
4263-
{
4264-
'type': 'output_text',
4265-
'text': reasoning_content,
4266-
}
4267-
]
4326+
reasoning_item['content'] = [{'type': 'output_text', 'text': reasoning_content}]
42684327

4328+
# Flush any buffered details-only chunks into this reasoning item.
4329+
if _pending_reasoning_details:
4330+
_merge_reasoning_details(
4331+
reasoning_item.setdefault('reasoning_details', []),
4332+
_pending_reasoning_details,
4333+
)
4334+
_pending_reasoning_details.clear()
4335+
4336+
# Accumulate raw structured reasoning_details for provider round-trip.
4337+
if reasoning_details_chunk:
4338+
if output and output[-1].get('type') == 'reasoning':
4339+
_merge_reasoning_details(
4340+
output[-1].setdefault('reasoning_details', []),
4341+
reasoning_details_chunk,
4342+
)
4343+
else:
4344+
# Buffer until a safe boundary (end-of-stream or real reasoning text).
4345+
items = reasoning_details_chunk if isinstance(reasoning_details_chunk, list) else [reasoning_details_chunk]
4346+
_pending_reasoning_details.extend(items)
4347+
4348+
if reasoning_content or reasoning_details_chunk:
42694349
data = {'content': serialize_output(full_output())}
42704350

42714351
if value:
@@ -4481,6 +4561,30 @@ async def flush_pending_delta_data(threshold: int = 0):
44814561
)
44824562
reasoning_item['status'] = 'completed'
44834563

4564+
# Flush any buffered reasoning_details that never found a reasoning item.
4565+
if _pending_reasoning_details:
4566+
target = next((item for item in output if item.get('type') == 'reasoning'), None)
4567+
if target is None:
4568+
target = {
4569+
'type': 'reasoning',
4570+
'id': output_id('r'),
4571+
'status': 'completed',
4572+
'start_tag': '<think>',
4573+
'end_tag': '</think>',
4574+
'attributes': {'type': 'reasoning_content'},
4575+
'content': [],
4576+
'summary': None,
4577+
'started_at': time.time(),
4578+
'ended_at': time.time(),
4579+
'duration': 0,
4580+
}
4581+
output.insert(0, target)
4582+
_merge_reasoning_details(
4583+
target.setdefault('reasoning_details', []),
4584+
_pending_reasoning_details,
4585+
)
4586+
_pending_reasoning_details.clear()
4587+
44844588
if response_tool_calls:
44854589
tool_calls.append(_split_tool_calls(response_tool_calls))
44864590

‎backend/open_webui/utils/misc.py‎

Lines changed: 26 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -250,10 +250,11 @@ def convert_output_to_messages(
250250
pending_tool_calls = []
251251
pending_content = []
252252
pending_reasoning = [] # Only populated when reasoning_format == 'reasoning_content'
253+
pending_reasoning_details = None
253254

254255
def flush_pending():
255-
nonlocal pending_content, pending_tool_calls, pending_reasoning
256-
if not pending_content and not pending_tool_calls and not pending_reasoning:
256+
nonlocal pending_content, pending_tool_calls, pending_reasoning, pending_reasoning_details
257+
if not pending_content and not pending_tool_calls and not pending_reasoning and not pending_reasoning_details:
257258
return
258259

259260
message = {
@@ -265,10 +266,14 @@ def flush_pending():
265266
if pending_reasoning:
266267
message['reasoning_content'] = '\n'.join(pending_reasoning)
267268

269+
if pending_reasoning_details:
270+
message['reasoning_details'] = pending_reasoning_details
271+
268272
messages.append(message)
269273
pending_content = []
270274
pending_tool_calls = []
271275
pending_reasoning = []
276+
pending_reasoning_details = None
272277

273278
for item in output:
274279
item_type = item.get('type', '')
@@ -339,7 +344,7 @@ def flush_pending():
339344
)
340345

341346
elif item_type == 'reasoning':
342-
if not reasoning_format:
347+
if not raw:
343348
continue
344349

345350
reasoning_text = ''
@@ -350,15 +355,28 @@ def flush_pending():
350355
elif 'text' in part:
351356
reasoning_text += part.get('text', '')
352357

358+
raw_details = item.get('reasoning_details')
359+
353360
if reasoning_text:
354-
if reasoning_format == 'think_tags':
355-
# Ollama: embed in content with the item's original tags
361+
if reasoning_format == 'reasoning_content':
362+
# llama.cpp: collect for reasoning_content field
363+
pending_reasoning.append(reasoning_text)
364+
elif not raw_details:
365+
# Wrap reasoning in <think> tags for Ollama and any
366+
# provider that lacks structured reasoning_details.
356367
start_tag = item.get('start_tag', '<think>')
357368
end_tag = item.get('end_tag', '</think>')
358369
pending_content.append(f'{start_tag}{reasoning_text}{end_tag}')
359-
elif reasoning_format == 'reasoning_content':
360-
# llama.cpp: collect for reasoning_content field
361-
pending_reasoning.append(reasoning_text)
370+
371+
# Preserve raw structured reasoning_details for provider round-trip
372+
if raw_details:
373+
if pending_reasoning_details is None:
374+
pending_reasoning_details = list(raw_details) if isinstance(raw_details, list) else [raw_details]
375+
elif isinstance(pending_reasoning_details, list):
376+
if isinstance(raw_details, list):
377+
pending_reasoning_details.extend(raw_details)
378+
else:
379+
pending_reasoning_details.append(raw_details)
362380

363381
elif item_type == 'open_webui:code_interpreter':
364382
# Always include code interpreter content so the LLM knows

0 commit comments

Comments
 (0)