my-project-public

repository

loading code, commits, and activity

repositories

loading repo index

#1	"""
#2	Mnemosyne Structured Fact Extraction
#3	====================================
#4	LLM-driven fact extraction as a derived layer.
#5	Extracts 2-5 concise factual statements from raw text.
#6	Facts are stored as TripleStore triples, not replacements for raw text.
#7
#8	Uses the same LLM fallback chain as local_llm.py:
#9
#10	0. Host-provided LLM backend (when MNEMOSYNE_HOST_LLM_ENABLED=true and a
#11	backend is registered). On host attempt with no usable output, skips
#12	the remote URL and goes straight to local GGUF.
#13	1. Remote OpenAI-compatible API (if MNEMOSYNE_LLM_BASE_URL set
#14	AND MNEMOSYNE_LLM_ENABLED is not false).
#15	2. Local ctransformers GGUF model.
#16	3. Skip extraction (graceful degradation).
#17
#18	Extraction uses temperature=0.0 (deterministic) so re-ingesting the same
#19	content does not create near-duplicate facts in the facts table.
#20	"""
#21
#22	import os
#23	from typing import List, Optional
#24
#25	# Reuse local_llm infrastructure
#26	from mnemosyne.core import local_llm
#27	from mnemosyne.core.local_llm import (
#28	llm_available,
#29	_call_remote_llm,
#30	_load_llm,
#31	_try_host_llm,
#32	LLM_BASE_URL,
#33	LLM_ENABLED,
#34	LLM_MAX_TOKENS,
#35	_clean_output,
#36	)
#37
#38	# --- Config ------------------------------------------------------------------
#39	EXTRACTION_PROMPT = os.environ.get(
#40	"MNEMOSYNE_EXTRACTION_PROMPT",
#41	"Extract 2-5 concise factual statements from the following text. "
#42	"Each fact should be a complete sentence describing something true about the subject. "
#43	"Focus on preferences, opinions, experiences, and factual claims. "
#44	"Return one fact per line. Do not number them. "
#45	"If no facts can be extracted, return 'NO_FACTS'.\n\nText: {text}\n\nFacts:"
#46	)
#47
#48
#49	def _build_extraction_prompt(text: str) -> str:
#50	"""Build the extraction prompt with the user text inserted."""
#51	return EXTRACTION_PROMPT.format(text=text)
#52
#53
#54	def _parse_facts(raw_output: str) -> List[str]:
#55	"""Parse LLM output into individual facts."""
#56	if not raw_output or raw_output.strip().upper() == "NO_FACTS":
#57	return []
#58
#59	# Split on newlines, filter empty lines
#60	lines = [line.strip() for line in raw_output.split("\n") if line.strip()]
#61
#62	# Clean up any numbering or bullet prefixes
#63	cleaned = []
#64	for line in lines:
#65	# Remove leading numbers/bullets: "1. fact" or "- fact" or "* fact"
#66	line = line.lstrip("0123456789.-* ").strip()
#67	if line and len(line) > 10: # Minimum fact length
#68	cleaned.append(line)
#69
#70	return cleaned[:5] # Cap at 5 facts
#71
#72
#73	def extract_facts(text: str) -> List[str]:
#74	"""
#75	Extract structured facts from raw text using LLM.
#76
#77	Args:
#78	text: Raw memory content to extract facts from
#79
#80	Returns:
#81	List of extracted fact strings (0-5 items). Empty list if LLM unavailable.
#82
#83	Notes:
#84	- The host backend (Hermes auxiliary client) is consulted first when
#85	enabled. Temperature is fixed at 0.0 so re-ingesting the same content
#86	produces deterministic facts (avoids near-duplicate writes to the
#87	facts table).
#88	- When the host attempt produces no usable text, the remote URL is
#89	skipped — falls through to local GGUF, then []. This honors the
#90	plan's host-vs-remote precedence rule.
#91	"""
#92	if not text or not text.strip():
#93	return []
#94
#95	if not local_llm.llm_available():
#96	return []
#97
#98	prompt = _build_extraction_prompt(text)
#99
#100	# 0. Host backend (deterministic; temperature=0.0).
#101	# Reference live module values so monkeypatch on local_llm reaches us.
#102	attempted, host_text = local_llm._try_host_llm(
#103	prompt, max_tokens=local_llm.LLM_MAX_TOKENS, temperature=0.0
#104	)
#105	if attempted:
#106	if host_text:
#107	facts = _parse_facts(host_text)
#108	if facts:
#109	return facts
#110	# Host attempted but produced no facts. Skip remote per A3; try local.
#111	llm = local_llm._load_llm()
#112	if llm is not None:
#113	try:
#114	raw_output = llm(
#115	prompt,
#116	max_new_tokens=local_llm.LLM_MAX_TOKENS,
#117	stop=["</s>", "<\|user\|>"],
#118	)
#119	return _parse_facts(local_llm._clean_output(raw_output))
#120	except Exception:
#121	return []
#122	return []
#123
#124	# 1. Remote LLM. Pass temperature=0.0 so the C2 determinism contract
#125	# holds even on the standalone remote path (where extract_facts shares
#126	# _call_remote_llm with summarize_memories' default of 0.3).
#127	if local_llm.LLM_ENABLED and local_llm.LLM_BASE_URL:
#128	raw_output = local_llm._call_remote_llm(prompt, temperature=0.0)
#129	if raw_output:
#130	facts = _parse_facts(local_llm._clean_output(raw_output))
#131	if facts:
#132	return facts
#133
#134	# 2. Local LLM.
#135	llm = local_llm._load_llm()
#136	if llm is not None:
#137	try:
#138	raw_output = llm(
#139	prompt,
#140	max_new_tokens=local_llm.LLM_MAX_TOKENS,
#141	stop=["</s>", "<\|user\|>"],
#142	)
#143	facts = _parse_facts(local_llm._clean_output(raw_output))
#144	return facts
#145	except Exception:
#146	pass
#147
#148	return []
#149
#150
#151	def extract_facts_safe(text: str) -> List[str]:
#152	"""
#153	Best-effort fact extraction that never raises.
#154	Wrapper for extract_facts with exception handling.
#155	"""
#156	try:
#157	return extract_facts(text)
#158	except Exception:
#159	return []
#160

z6Mkq5mY3JWtxoxUobWcfNHm7AkRubgSWEZTkBVqZXJviFZ5/my-project-public