admin
7a31be9de5
- Configurable summary prompts via ENV or file; stronger default prompt. - New docx_styles.py: converts markdown (headings, bullets, bold/italic) to DOCX. - Updated create_summary_docx to use markdown-aware styling. - New docx_cover.py: reusable cover page for transcript and summary. - Cover pages enabled when COVER_PAGE_ENABLED=true.
295 lines
10 KiB
Python
295 lines
10 KiB
Python
"""
|
|
Summarizer Module
|
|
-----------------
|
|
|
|
Provides a client to summarize long transcripts via an LLM endpoint.
|
|
|
|
Behavior:
|
|
- Chunks transcript into 10,240-character segments.
|
|
- Summarizes each chunk.
|
|
- Summarizes the summaries into a final, detailed summary.
|
|
|
|
Environment Variables:
|
|
- SUMMARIZER_API_URL: (required) Base URL of the LLM API (e.g., http://localhost:8080)
|
|
- SUMMARIZER_API_KEY: (optional) API key, if required
|
|
- SUMMARIZER_MODEL: (optional) Model name (e.g., llama-3.1-8b-instruct)
|
|
"""
|
|
|
|
import os
|
|
import json
|
|
import logging
|
|
from typing import Optional
|
|
|
|
import httpx
|
|
|
|
logger = logging.getLogger("scraibe.summarizer")
|
|
|
|
|
|
class SummarizerError(Exception):
|
|
"""Raised when the summarization API call fails."""
|
|
pass
|
|
|
|
|
|
class SummarizerClient:
|
|
"""
|
|
HTTP client for an OpenAI-compatible chat completions endpoint.
|
|
Used to summarize long transcripts in chunks.
|
|
"""
|
|
|
|
CHUNK_SIZE = 10_240 # characters per chunk
|
|
|
|
def __init__(
|
|
self,
|
|
api_url: Optional[str] = None,
|
|
api_key: Optional[str] = None,
|
|
model: Optional[str] = None,
|
|
timeout: float = 3600.0,
|
|
):
|
|
self.api_url = (api_url or os.getenv("SUMMARIZER_API_URL")).strip().rstrip("/")
|
|
self.api_key = api_key or os.getenv("SUMMARIZER_API_KEY") or None
|
|
self.model = model or os.getenv("SUMMARIZER_MODEL") or "llama-3.1-8b-instruct"
|
|
self.timeout = timeout
|
|
|
|
if not self.api_url:
|
|
raise SummarizerError(
|
|
"SUMMARIZER_API_URL is not set. "
|
|
"Provide the summarization LLM URL via environment or constructor."
|
|
)
|
|
|
|
logger.info(
|
|
"Initializing SummarizerClient: url=%s model=%s",
|
|
self.api_url,
|
|
self.model,
|
|
)
|
|
|
|
self._client = httpx.Client(
|
|
base_url=self.api_url,
|
|
timeout=self.timeout,
|
|
follow_redirects=True,
|
|
)
|
|
|
|
def close(self):
|
|
self._client.close()
|
|
|
|
def __del__(self):
|
|
try:
|
|
self._client.close()
|
|
except Exception:
|
|
pass
|
|
|
|
def summarize_transcript(self, transcript: str) -> str:
|
|
"""
|
|
Summarize a (possibly very long) transcript.
|
|
|
|
Strategy:
|
|
- Split transcript into chunks of CHUNK_SIZE characters.
|
|
- Generate a detailed summary for each chunk.
|
|
- Combine all chunk summaries and generate a final, concise but thorough summary.
|
|
|
|
The final summary should make it clear:
|
|
- What was discussed
|
|
- Main issues
|
|
- Outcomes / decisions
|
|
- Next steps / action items
|
|
"""
|
|
if not transcript.strip():
|
|
logger.warning("Empty transcript provided to summarize_transcript.")
|
|
return "No transcript provided to summarize."
|
|
|
|
logger.info(
|
|
"Starting summarization for transcript length=%d chars",
|
|
len(transcript),
|
|
)
|
|
|
|
# 1) Chunk the transcript
|
|
chunks = self._chunk_text(transcript)
|
|
logger.info("Split transcript into %d chunks.", len(chunks))
|
|
|
|
# 2) Summarize each chunk
|
|
chunk_summaries = []
|
|
for i, chunk in enumerate(chunks):
|
|
logger.info(
|
|
"Summarizing chunk %d/%d (length=%d)",
|
|
i + 1,
|
|
len(chunks),
|
|
len(chunk),
|
|
)
|
|
summary = self._summarize_chunk(chunk, i, len(chunks))
|
|
chunk_summaries.append(summary)
|
|
|
|
# 3) Combine and summarize summaries
|
|
combined = "\n\n".join(chunk_summaries)
|
|
logger.info(
|
|
"Combining %d chunk summaries (total length=%d) for final summary.",
|
|
len(chunk_summaries),
|
|
len(combined),
|
|
)
|
|
final_summary = self._summarize_combined(combined)
|
|
|
|
logger.info("Summarization completed.")
|
|
return final_summary
|
|
|
|
def _chunk_text(self, text: str) -> list[str]:
|
|
"""Split text into chunks of CHUNK_SIZE characters."""
|
|
chunks = []
|
|
start = 0
|
|
while start < len(text):
|
|
end = start + self.CHUNK_SIZE
|
|
if end >= len(text):
|
|
chunks.append(text[start:])
|
|
break
|
|
# Try to break at a reasonable boundary (newline or space)
|
|
break_pos = text.rfind("\n", start, end)
|
|
if break_pos == -1:
|
|
break_pos = text.rfind(" ", start, end)
|
|
if break_pos == -1 or break_pos <= start:
|
|
break_pos = end
|
|
chunks.append(text[start:break_pos].strip())
|
|
start = break_pos
|
|
return chunks
|
|
|
|
def _load_summary_prompt(self, role: str) -> str:
|
|
"""
|
|
Load summary prompt for the given role: 'chunk' or 'combined'.
|
|
|
|
Priority:
|
|
1) SUMMARY_PROMPT_{ROLE} (env)
|
|
2) SUMMARY_PROMPT_FILE (env) with [chunk] / [combined] sections
|
|
3) Built-in default prompt
|
|
"""
|
|
role_upper = role.upper()
|
|
|
|
# 1) Direct env var: SUMMARY_PROMPT_CHUNK / SUMMARY_PROMPT_COMBINED
|
|
env_key = f"SUMMARY_PROMPT_{role_upper}"
|
|
env_prompt = (os.getenv(env_key) or "").strip()
|
|
if env_prompt:
|
|
return env_prompt
|
|
|
|
# 2) File-based prompt with sections
|
|
prompt_file = (os.getenv("SUMMARY_PROMPT_FILE") or "").strip()
|
|
if prompt_file and os.path.exists(prompt_file):
|
|
try:
|
|
with open(prompt_file, "r", encoding="utf-8") as f:
|
|
content = f.read()
|
|
# Simple section parser: [chunk], [combined]
|
|
import re
|
|
pattern = re.compile(
|
|
r"\[" + role + r"\]\s*\n(.*?)(?=\n\[|$)",
|
|
re.DOTALL,
|
|
)
|
|
m = pattern.search(content)
|
|
if m:
|
|
text = m.group(1).strip()
|
|
if text:
|
|
return text
|
|
except Exception as e:
|
|
logger.warning("Failed to load SUMMARY_PROMPT_FILE for %s: %s", role, e)
|
|
|
|
# 3) Default prompts
|
|
if role == "chunk":
|
|
return (
|
|
"You are an expert legal and business meeting summarizer. "
|
|
"You will receive a segment of a longer transcript. "
|
|
"Provide a detailed, structured summary of this segment, focusing on: "
|
|
"- Topics discussed\n"
|
|
"- Key points and arguments\n"
|
|
"- Decisions and agreements\n"
|
|
"- Action items and responsibilities\n"
|
|
"- Any risks, conflicts, or open issues\n\n"
|
|
"Be concise but complete. Use bullet points where helpful. "
|
|
"Do not add information that is not present in the transcript."
|
|
)
|
|
else:
|
|
return (
|
|
"You are an expert legal and business meeting summarizer. "
|
|
"You will receive several intermediate summaries of a longer conversation. "
|
|
"Produce a single, comprehensive summary that makes it clear: "
|
|
"- The overall purpose and context of the discussion\n"
|
|
"- The main issues and topics addressed\n"
|
|
"- Key arguments and positions (briefly)\n"
|
|
"- Decisions and outcomes\n"
|
|
"- Action items, responsibilities, and next steps\n"
|
|
"- Any unresolved issues or risks\n\n"
|
|
"The summary should be detailed enough that a reader who was not present "
|
|
"can understand what happened and what is expected going forward. "
|
|
"Use clear, concise language and bullet points where appropriate. "
|
|
"Use markdown formatting (headings, lists, bold) to structure the summary."
|
|
)
|
|
|
|
def _summarize_chunk(self, chunk: str, index: int, total: int) -> str:
|
|
system_prompt = self._load_summary_prompt("chunk")
|
|
|
|
user_prompt = (
|
|
f"This is segment {index + 1} of {total} from a longer conversation.\n\n"
|
|
f"{chunk}"
|
|
)
|
|
|
|
return self._chat_completion(system_prompt, user_prompt)
|
|
|
|
def _summarize_combined(self, combined_summaries: str) -> str:
|
|
system_prompt = self._load_summary_prompt("combined")
|
|
|
|
user_prompt = (
|
|
"Here are the intermediate summaries from different parts of the same conversation:\n\n"
|
|
f"{combined_summaries}"
|
|
)
|
|
|
|
return self._chat_completion(system_prompt, user_prompt)
|
|
|
|
def _chat_completion(self, system_prompt: str, user_prompt: str) -> str:
|
|
"""
|
|
Call OpenAI-compatible /v1/chat/completions endpoint.
|
|
"""
|
|
payload = {
|
|
"model": self.model,
|
|
"messages": [
|
|
{"role": "system", "content": system_prompt},
|
|
{"role": "user", "content": user_prompt},
|
|
],
|
|
"temperature": 0.3,
|
|
}
|
|
|
|
headers = {
|
|
"Content-Type": "application/json",
|
|
}
|
|
if self.api_key:
|
|
headers["Authorization"] = f"Bearer {self.api_key}"
|
|
|
|
logger.info("Calling summarizer endpoint: /v1/chat/completions")
|
|
|
|
resp = self._client.post(
|
|
"/v1/chat/completions",
|
|
json=payload,
|
|
headers=headers,
|
|
)
|
|
|
|
logger.info("Summarizer response status: %d", resp.status_code)
|
|
|
|
if resp.status_code >= 400:
|
|
logger.error("Summarizer error response: %s", resp.text)
|
|
raise SummarizerError(
|
|
f"Summarizer API error {resp.status_code}: {resp.text}"
|
|
)
|
|
|
|
try:
|
|
data = resp.json()
|
|
except json.JSONDecodeError:
|
|
logger.error("Failed to parse summarizer response as JSON.")
|
|
raise SummarizerError(
|
|
"Failed to parse summarizer response as JSON."
|
|
)
|
|
|
|
# Extract assistant message
|
|
try:
|
|
content = data["choices"][0]["message"]["content"]
|
|
return content.strip()
|
|
except (KeyError, IndexError, TypeError):
|
|
logger.error(
|
|
"Unexpected summarizer response format: %s",
|
|
json.dumps(data, indent=2),
|
|
)
|
|
raise SummarizerError(
|
|
"Unexpected summarizer response format: "
|
|
f"{json.dumps(data, indent=2)}"
|
|
)
|