fleet-memory/memory/llm_client.py
Nicolò Boschi bc7d9fe07f improvements
2025-10-30 18:54:06 +01:00

294 lines
11 KiB
Python

"""
LLM client for fact extraction and other AI-powered operations.
Uses OpenAI-compatible API (works with Groq, OpenAI, etc.)
"""
import os
import json
import re
import asyncio
from datetime import datetime
from typing import List, Dict, Optional, Literal
from openai import AsyncOpenAI
from pydantic import BaseModel, Field
class ExtractedFact(BaseModel):
"""A single extracted fact from text."""
fact: str = Field(
description="Self-contained factual statement with subject + action + context"
)
date: str = Field(
description="Absolute date/time when this fact occurred in ISO format (YYYY-MM-DDTHH:MM:SSZ). If text mentions relative time (yesterday, last week, this morning), calculate absolute date from the provided context date."
)
class FactExtractionResponse(BaseModel):
"""Response containing all extracted facts."""
facts: List[ExtractedFact] = Field(
description="List of extracted factual statements"
)
def chunk_text(text: str, max_chars: int = 120000) -> List[str]:
"""
Split text into chunks at sentence boundaries using LangChain's text splitter.
Uses RecursiveCharacterTextSplitter which intelligently splits at sentence boundaries
and allows chunks to slightly exceed max_chars to finish sentences naturally.
Args:
text: Input text to chunk
max_chars: Maximum characters per chunk (default 120k ≈ 30k tokens)
Note: chunks may slightly exceed this to complete sentences
Returns:
List of text chunks, roughly under max_chars
"""
from langchain_text_splitters import RecursiveCharacterTextSplitter
# If text is small enough, return as-is
if len(text) <= max_chars:
return [text]
# Configure splitter to split at sentence boundaries first
# Separators in order of preference: paragraphs, newlines, sentences, words
splitter = RecursiveCharacterTextSplitter(
chunk_size=max_chars,
chunk_overlap=0,
length_function=len,
is_separator_regex=False,
separators=[
"\n\n", # Paragraph breaks
"\n", # Line breaks
". ", # Sentence endings
"! ", # Exclamations
"? ", # Questions
"; ", # Semicolons
", ", # Commas
" ", # Words
"", # Characters (last resort)
],
)
return splitter.split_text(text)
def get_llm_client() -> AsyncOpenAI:
"""
Get configured async LLM client.
Supports:
- Groq (default): Set GROQ_API_KEY and optionally GROQ_BASE_URL
- OpenAI: Set OPENAI_API_KEY
Returns:
Configured AsyncOpenAI client
"""
# Check for Groq configuration first
groq_api_key = os.getenv('GROQ_API_KEY')
if groq_api_key:
base_url = os.getenv('GROQ_BASE_URL', 'https://api.groq.com/openai/v1')
return AsyncOpenAI(
api_key=groq_api_key,
base_url=base_url
)
# Fall back to OpenAI
openai_api_key = os.getenv('OPENAI_API_KEY')
if openai_api_key:
return AsyncOpenAI(api_key=openai_api_key)
raise ValueError(
"No LLM API key found. Set GROQ_API_KEY or OPENAI_API_KEY environment variable."
)
async def _extract_facts_from_chunk(
chunk: str,
chunk_index: int,
total_chunks: int,
event_date: datetime,
context: str,
model: str,
temperature: float,
max_tokens: int,
client: AsyncOpenAI
) -> List[Dict[str, str]]:
"""
Extract facts from a single chunk (internal helper for parallel processing).
"""
# Format event_date for the prompt
event_date_str = event_date.strftime("%Y-%m-%dT%H:%M:%SZ")
prompt = f"""You are extracting facts from text for an AI memory system. Each fact will be stored and retrieved later.
## CONTEXT INFORMATION
- Current reference date/time: {event_date_str}
- Context: {context if context else 'no context provided'}
## CRITICAL: Facts must be DETAILED and COMPREHENSIVE
Each fact should:
1. Be SELF-CONTAINED - readable without the original context
2. Include ALL relevant details: WHO, WHAT, WHERE, WHEN, WHY, HOW
3. Preserve specific names, dates, numbers, locations, relationships
4. Resolve pronouns to actual names/entities
5. Include surrounding context that makes the fact meaningful
6. Capture nuances, reasons, causes, and implications
## TEMPORAL INFORMATION (VERY IMPORTANT)
For each fact, extract the ABSOLUTE date/time when it occurred:
- If text mentions ABSOLUTE dates ("on March 15, 2024", "last Tuesday"), use that date
- If text mentions RELATIVE times ("yesterday", "last week", "this morning", "3 days ago"), calculate the absolute date using the reference date above.
- if text mentions a vague relative time without a specific day ("last week", "this morning"), transform the date in relative with absolute context ("last week" + " 2 june 2024" -> "week before June 2 2024") in the text and use the absolute date for the 'date' field
- If NO specific time is mentioned, use the reference date
- Always output dates in ISO format: YYYY-MM-DDTHH:MM:SSZ
Examples of date extraction:
- Reference: 2024-03-20T10:00:00Z
- "Yesterday I went hiking" → date: 2024-03-19T10:00:00Z
- "Last week I joined Google" → date: 2024-03-13T10:00:00Z (approximately)
- "This morning I had coffee" → date: 2024-03-20T08:00:00Z
- "I work at Google" (no time mentioned) → date: 2024-03-20T10:00:00Z (use reference)
## What to EXTRACT (BE EXHAUSTIVE - DO NOT SKIP ANYTHING):
- **Biographical information**: jobs, roles, backgrounds, experiences, skills
- **Events (NEVER MISS THESE)**:
- ANY action that happened (went, did, attended, joined, started, finished, etc.)
- Photos, images, videos shared or taken ("here's a photo", "took a picture", "captured")
- Social activities (meetups, gatherings, meals, conversations)
- Achievements, milestones, accomplishments
- Travels, visits, locations visited
- Purchases, acquisitions, creations
- **Opinions and beliefs**: who believes what and why
- **Recommendations and advice**: specific suggestions with reasoning
- **Descriptions**: detailed explanations of how things work
- **Relationships**: connections between people, organizations, concepts
- **States and conditions**: current status, ongoing situations
## CRITICAL: Extract EVERY event mentioned, even casual ones
- "here's a photo of X" = someone took/shared a photo of X
- "I was with friends last week" = meetup/gathering with friends last week
- "sent you that link" = action of sending a link
- DO NOT skip events just because they seem minor or casual
## What to SKIP (ONLY these):
- Greetings, thank yous, acknowledgments (unless they reveal information)
- Filler words ("um", "uh", "like")
- Pure reactions without content ("wow", "cool", "nice")
- Incomplete thoughts or sentence fragments with no meaning
## EXAMPLES of GOOD facts (detailed, comprehensive):
Input: "Alice mentioned she works at Google in Mountain View. She joined the AI team last year."
GOOD fact: "Alice works at Google in Mountain View on the AI team, which she joined last year"
GOOD date: Calculate based on reference date (if reference is 2024-03-20, "last year" = 2023-03-20)
Input: "Yesterday Bob went hiking in Yosemite because it helps him clear his mind."
GOOD fact: "Bob went hiking in Yosemite because it helps him clear his mind"
GOOD date: Reference date minus 1 day
Input: "Here's a photo of me with my friends taken last week at the beach."
GOOD fact: "Someone shared/took a photo with their friends at the beach"
GOOD date: Reference date minus 7 days (last week)
NOTE: Extract the event (photo taken/shared with friends at beach), NOT just that a photo exists
Input: "I sent you that article about AI last Tuesday."
GOOD fact: "Someone sent an article about AI"
GOOD date: Calculate last Tuesday from reference date
## TEXT TO EXTRACT FROM:
{chunk}
Remember:
1. BE EXHAUSTIVE - Extract EVERY event, action, and fact mentioned
2. DO NOT skip casual mentions like "here's a photo", "I was with X", "sent you Y"
3. Include ALL details, names, numbers, reasons, and context in the fact text
4. Extract the absolute date for EACH fact by calculating relative times from the reference date
5. When in doubt, EXTRACT IT - better to have too many facts than miss important events"""
response = await client.beta.chat.completions.parse(
model=model,
messages=[
{
"role": "system",
"content": "You are an EXHAUSTIVE fact extractor. Extract EVERY event, action, and fact mentioned - never skip anything. This includes casual mentions like photos shared, things sent, meetups, gatherings, or any action. Preserve all context, details, and nuances. Calculate absolute dates from relative time expressions. When in doubt, extract it - missing facts is worse than extracting too many."
},
{
"role": "user",
"content": prompt
}
],
temperature=temperature,
max_tokens=max_tokens,
response_format=FactExtractionResponse,
extra_body={"service_tier": "auto"},
)
# Extract the parsed response
extraction_response = response.choices[0].message.parsed
# Convert to dict format
chunk_facts = [fact.model_dump() for fact in extraction_response.facts]
return chunk_facts
async def extract_facts_from_text(
text: str,
event_date: datetime,
context: str = "",
model: str = "openai/gpt-oss-120b",
temperature: float = 0.1,
max_tokens: int = 65000,
chunk_size: int = 5000
) -> List[Dict[str, str]]:
"""
Extract semantic facts from conversational or narrative text using LLM.
For large texts (>chunk_size chars), automatically chunks at sentence boundaries
to avoid hitting output token limits. Processes ALL chunks in PARALLEL for speed.
Args:
text: Input text (conversation, article, etc.)
event_date: Reference date for resolving relative times
context: Context about the conversation/document
model: LLM model to use
temperature: Sampling temperature (lower = more focused)
max_tokens: Maximum tokens in response
chunk_size: Maximum characters per chunk
Returns:
List of fact dictionaries with 'fact' and 'date' keys
"""
client = get_llm_client()
# Chunk text if necessary
chunks = chunk_text(text, max_chars=chunk_size)
# Process all chunks in parallel using asyncio.gather
tasks = [
_extract_facts_from_chunk(
chunk=chunk,
chunk_index=i,
total_chunks=len(chunks),
event_date=event_date,
context=context,
model=model,
temperature=temperature,
max_tokens=max_tokens,
client=client
)
for i, chunk in enumerate(chunks)
]
# Wait for all chunks to complete in parallel
chunk_results = await asyncio.gather(*tasks)
# Flatten results from all chunks
all_facts = []
for chunk_facts in chunk_results:
all_facts.extend(chunk_facts)
return all_facts