selfhostyourtech/reflector
2
0
Fork 0
mirror of https://github.com/Monadical-SAS/reflector.git synced 2025-12-22 05:09:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat: use llamaindex everywhere (#525)

Browse Source 
* feat: use llamaindex for transcript final title too

* refactor: removed llm backend, replaced with one single class+llamaindex

* refactor: self-review

* fix: typing

* fix: tests

* refactor: extract clean_title and add tests

* test: fix

* test: remove ensure_casing/nltk

* fix: tiny mistake
This commit is contained in:
Mathieu Virbel
2025-08-01 12:13:00 -06:00
committed by GitHub
parent 1878834ce6
commit 28ac031ff6
25 changed files with 284 additions and 1539 deletions
Download Patch File Download Diff File Expand all files Collapse all files

63
server/reflector/processors/summary/summary_builder.py
View File

@@ -12,15 +12,9 @@ from textwrap import dedent
from typing import Type, TypeVar
import structlog
from llama_index.core import Settings
from llama_index.core.output_parsers import PydanticOutputParser
from llama_index.core.program import LLMTextCompletionProgram
from llama_index.core.response_synthesizers import TreeSummarize
from llama_index.llms.openai_like import OpenAILike
from pydantic import BaseModel, Field
from reflector.llm.base import LLM
from reflector.llm.openai_llm import OpenAILLM
from reflector.llm import LLM
from reflector.settings import settings
T = TypeVar("T", bound=BaseModel)
@@ -168,23 +162,12 @@ class SummaryBuilder:
self.summaries: list[dict[str, str]] = []
self.subjects: list[str] = []
self.transcription_type: TranscriptionType | None = None
self.llm_instance: LLM = llm
self.llm: LLM = llm
self.model_name: str = llm.model_name
self.logger = logger or structlog.get_logger()
if filename:
self.read_transcript_from_file(filename)
Settings.llm = OpenAILike(
model=llm.model_name,
api_base=llm.url,
api_key=llm.api_key,
context_window=settings.SUMMARY_LLM_CONTEXT_SIZE_TOKENS,
is_chat_model=True,
is_function_calling_model=llm.has_structured_output,
temperature=llm.temperature,
max_tokens=llm.max_tokens,
)
def read_transcript_from_file(self, filename: str) -> None:
"""
Load a transcript from a text file.
@@ -202,40 +185,16 @@ class SummaryBuilder:
self.transcript = transcript
def set_llm_instance(self, llm: LLM) -> None:
self.llm_instance = llm
self.llm = llm
async def _get_structured_response(
self, prompt: str, output_cls: Type[T], tone_name: str | None = None
) -> Type[T]:
) -> T:
"""Generic function to get structured output from LLM for non-function-calling models."""
# First, use TreeSummarize to get the response
summarizer = TreeSummarize(verbose=True)
response = await summarizer.aget_response(
prompt, [self.transcript], tone_name=tone_name
return await self.llm.get_structured_response(
prompt, [self.transcript], output_cls, tone_name=tone_name
)
# Then, use PydanticOutputParser to structure the response
output_parser = PydanticOutputParser(output_cls)
prompt_template_str = STRUCTURED_RESPONSE_PROMPT_TEMPLATE
program = LLMTextCompletionProgram.from_defaults(
output_parser=output_parser,
prompt_template_str=prompt_template_str,
verbose=False,
)
format_instructions = output_parser.format(
"Please structure the above information in the following JSON format:"
)
output = await program.acall(
analysis=str(response), format_instructions=format_instructions
)
return output
# ----------------------------------------------------------------------------
# Participants
# ----------------------------------------------------------------------------
@@ -354,19 +313,18 @@ class SummaryBuilder:
async def generate_subject_summaries(self) -> None:
"""Generate detailed summaries for each extracted subject."""
assert self.transcript is not None
summarizer = TreeSummarize(verbose=False)
summaries = []
for subject in self.subjects:
detailed_prompt = DETAILED_SUBJECT_PROMPT_TEMPLATE.format(subject=subject)
detailed_response = await summarizer.aget_response(
detailed_response = await self.llm.get_response(
detailed_prompt, [self.transcript], tone_name="Topic assistant"
)
paragraph_prompt = PARAGRAPH_SUMMARY_PROMPT
paragraph_response = await summarizer.aget_response(
paragraph_response = await self.llm.get_response(
paragraph_prompt, [str(detailed_response)], tone_name="Topic summarizer"
)
@@ -377,7 +335,6 @@ class SummaryBuilder:
async def generate_recap(self) -> None:
"""Generate a quick recap from the subject summaries."""
summarizer = TreeSummarize(verbose=True)
summaries_text = "\n\n".join(
[
@@ -388,7 +345,7 @@ class SummaryBuilder:
recap_prompt = RECAP_PROMPT
recap_response = await summarizer.aget_response(
recap_response = await self.llm.get_response(
recap_prompt, [summaries_text], tone_name="Recap summarizer"
)
@@ -483,7 +440,7 @@ if __name__ == "__main__":
async def main():
# build the summary
llm = OpenAILLM(config_prefix="SUMMARY", settings=settings)
llm = LLM(settings=settings)
sm = SummaryBuilder(llm=llm, filename=args.transcript)
if args.subjects:

4
server/reflector/processors/transcript_final_summary.py
View File

@@ -1,4 +1,4 @@
from reflector.llm.openai_llm import OpenAILLM
from reflector.llm import LLM
from reflector.processors.base import Processor
from reflector.processors.summary.summary_builder import SummaryBuilder
from reflector.processors.types import FinalLongSummary, FinalShortSummary, TitleSummary
@@ -17,7 +17,7 @@ class TranscriptFinalSummaryProcessor(Processor):
super().__init__(**kwargs)
self.transcript = transcript
self.chunks: list[TitleSummary] = []
self.llm = OpenAILLM(config_prefix="SUMMARY", settings=settings)
self.llm = LLM(settings=settings)
self.builder = None
async def _push(self, data: TitleSummary):

79
server/reflector/processors/transcript_final_title.py
View File

@@ -1,67 +1,72 @@
from reflector.llm import LLM, LLMTaskParams
from textwrap import dedent
from reflector.llm import LLM
from reflector.processors.base import Processor
from reflector.processors.types import FinalTitle, TitleSummary
from reflector.settings import settings
from reflector.utils.text import clean_title
TITLE_PROMPT = dedent(
"""
Generate a concise title for this meeting based on the following topic titles.
Ignore casual conversation, greetings, or administrative matters.
The title must:
- Be maximum 10 words
- Use noun phrases when possible (e.g., "Q1 Budget Review" not "Reviewing the Q1 Budget")
- Avoid generic terms like "Team Meeting" or "Discussion"
If multiple unrelated topics were discussed, prioritize the most significant one.
or create a compound title (e.g., "Product Launch and Budget Planning").
<topics_discussed>
{titles}
</topics_discussed>
Do not explain, just output the meeting title as a single line.
"""
).strip()
class TranscriptFinalTitleProcessor(Processor):
"""
Assemble all summary into a line-based json
Generate a final title from topic titles using LlamaIndex
"""
INPUT_TYPE = TitleSummary
OUTPUT_TYPE = FinalTitle
TASK = "final_title"
def __init__(self, **kwargs):
super().__init__(**kwargs)
self.chunks: list[TitleSummary] = []
self.llm = LLM.get_instance()
self.params = LLMTaskParams.get_instance(self.TASK).task_params
self.llm = LLM(settings=settings, temperature=0.5, max_tokens=200)
async def _push(self, data: TitleSummary):
self.chunks.append(data)
async def get_title(self, text: str) -> dict:
async def get_title(self, accumulated_titles: str) -> str:
"""
Generate a title for the whole recording
Generate a title for the whole recording using LLM
"""
chunks = list(self.llm.split_corpus(corpus=text, task_params=self.params))
prompt = TITLE_PROMPT.format(titles=accumulated_titles)
response = await self.llm.get_response(
prompt,
[accumulated_titles],
tone_name="Title generator",
)
if len(chunks) == 1:
chunk = chunks[0]
prompt = self.llm.create_prompt(instruct=self.params.instruct, text=chunk)
title_result = await self.llm.generate(
prompt=prompt,
gen_schema=self.params.gen_schema,
gen_cfg=self.params.gen_cfg,
logger=self.logger,
)
return title_result
else:
accumulated_titles = ""
for chunk in chunks:
prompt = self.llm.create_prompt(
instruct=self.params.instruct, text=chunk
)
title_result = await self.llm.generate(
prompt=prompt,
gen_schema=self.params.gen_schema,
gen_cfg=self.params.gen_cfg,
logger=self.logger,
)
accumulated_titles += title_result["title"]
self.logger.info(f"Generated title response: {response}")
return await self.get_title(accumulated_titles)
return response
async def _flush(self):
if not self.chunks:
self.logger.warning("No summary to output")
return
accumulated_titles = ".".join([chunk.title for chunk in self.chunks])
title_result = await self.get_title(accumulated_titles)
final_title = self.llm.trim_title(title_result["title"])
final_title = self.llm.ensure_casing(final_title)
accumulated_titles = "\n".join([f"- {chunk.title}" for chunk in self.chunks])
title = await self.get_title(accumulated_titles)
title = clean_title(title)
final_title = FinalTitle(title=final_title)
final_title = FinalTitle(title=title)
await self.emit(final_title)

61
server/reflector/processors/transcript_topic_detector.py
View File

@@ -1,7 +1,41 @@
from reflector.llm import LLM, LLMTaskParams
from textwrap import dedent
from pydantic import BaseModel, Field
from reflector.llm import LLM
from reflector.processors.base import Processor
from reflector.processors.types import TitleSummary, Transcript
from reflector.settings import settings
from reflector.utils.text import clean_title
TOPIC_PROMPT = dedent(
"""
Analyze the following transcript segment and extract the main topic being discussed.
Focus on the substantive content and ignore small talk or administrative chatter.
Create a title that:
- Captures the specific subject matter being discussed
- Is descriptive and self-explanatory
- Uses professional language
- Is specific rather than generic
For the summary:
- Summarize the key points in maximum two sentences
- Focus on what was discussed, decided, or accomplished
- Be concise but informative
<transcript>
{text}
</transcript>
"""
).strip()
class TopicResponse(BaseModel):
"""Structured response for topic detection"""
title: str = Field(description="A descriptive title for the topic being discussed")
summary: str = Field(description="A concise 1-2 sentence summary of the discussion")
class TranscriptTopicDetectorProcessor(Processor):
@@ -11,7 +45,6 @@ class TranscriptTopicDetectorProcessor(Processor):
INPUT_TYPE = Transcript
OUTPUT_TYPE = TitleSummary
TASK = "topic"
def __init__(
self, min_transcript_length: int = int(settings.MIN_TRANSCRIPT_LENGTH), **kwargs
@@ -19,8 +52,7 @@ class TranscriptTopicDetectorProcessor(Processor):
super().__init__(**kwargs)
self.transcript = None
self.min_transcript_length = min_transcript_length
self.llm = LLM.get_instance()
self.params = LLMTaskParams.get_instance(self.TASK).task_params
self.llm = LLM(settings=settings, temperature=0.9, max_tokens=500)
async def _push(self, data: Transcript):
if self.transcript is None:
@@ -34,18 +66,15 @@ class TranscriptTopicDetectorProcessor(Processor):
return
await self.flush()
async def get_topic(self, text: str) -> dict:
async def get_topic(self, text: str) -> TopicResponse:
"""
Generate a topic and description for a transcription excerpt
Generate a topic and description for a transcription excerpt using LLM
"""
prompt = self.llm.create_prompt(instruct=self.params.instruct, text=text)
topic_result = await self.llm.generate(
prompt=prompt,
gen_schema=self.params.gen_schema,
gen_cfg=self.params.gen_cfg,
logger=self.logger,
prompt = TOPIC_PROMPT.format(text=text)
response = await self.llm.get_structured_response(
prompt, [text], TopicResponse, tone_name="Topic analyzer"
)
return topic_result
return response
async def _flush(self):
if not self.transcript:
@@ -53,13 +82,13 @@ class TranscriptTopicDetectorProcessor(Processor):
text = self.transcript.text
self.logger.info(f"Topic detector got {len(text)} length transcript")
topic_result = await self.get_topic(text=text)
title = self.llm.trim_title(topic_result["title"])
title = self.llm.ensure_casing(title)
title = clean_title(topic_result.title)
summary = TitleSummary(
title=title,
summary=topic_result["summary"],
summary=topic_result.summary,
timestamp=self.transcript.timestamp,
duration=self.transcript.duration,
transcript=self.transcript,

3
server/reflector/processors/transcript_translator.py
View File

@@ -13,14 +13,13 @@ class TranscriptTranslatorProcessor(Processor):
INPUT_TYPE = Transcript
OUTPUT_TYPE = Transcript
TASK = "translate"
def __init__(self, **kwargs):
super().__init__(**kwargs)
self.transcript = None
self.translate_url = settings.TRANSLATE_URL
self.timeout = settings.TRANSLATE_TIMEOUT
self.headers = {"Authorization": f"Bearer {settings.LLM_MODAL_API_KEY}"}
self.headers = {"Authorization": f"Bearer {settings.TRANSCRIPT_MODAL_API_KEY}"}
async def _push(self, data: Transcript):
self.transcript = data