feat: structured user profile with LLM-learned preferences

Replaces the freeform briefing-profile note with a DB-backed user_profiles
table. Users can edit job/industry/expertise/response preferences/interests/
work schedule via a new Settings → Profile tab. The LLM appends nightly
observations; at 14+ entries they are auto-consolidated into a learned_summary.
Profile context is injected into both briefing and chat system prompts.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
2026-03-30 14:17:30 -04:00
parent 9f3b9e45c6
commit dba41879ed
11 changed files with 667 additions and 9 deletions
@@ -0,0 +1,203 @@
"""User profile service — structured per-user preferences for LLM context."""
import asyncio
import logging
from datetime import date, datetime, timedelta, timezone
from sqlalchemy import select
from fabledassistant.models import async_session
from fabledassistant.models.user_profile import UserProfile
logger = logging.getLogger(__name__)
VALID_EXPERTISE = {"novice", "intermediate", "expert"}
VALID_STYLES = {"concise", "balanced", "detailed"}
VALID_TONES = {"casual", "professional", "technical"}
# Trigger consolidation when raw observations reach this count
_CONSOLIDATION_THRESHOLD = 14
async def get_profile(user_id: int) -> UserProfile:
"""Get or create the profile row for a user."""
async with async_session() as session:
result = await session.execute(
select(UserProfile).where(UserProfile.user_id == user_id)
)
profile = result.scalar_one_or_none()
if profile is None:
profile = UserProfile(user_id=user_id)
session.add(profile)
await session.commit()
await session.refresh(profile)
return profile
async def update_profile(user_id: int, data: dict) -> UserProfile:
"""Upsert structured profile fields from a validated dict."""
allowed = {
"display_name", "job_title", "industry", "expertise_level",
"response_style", "tone", "interests", "work_schedule",
}
async with async_session() as session:
result = await session.execute(
select(UserProfile).where(UserProfile.user_id == user_id)
)
profile = result.scalar_one_or_none()
if profile is None:
profile = UserProfile(user_id=user_id)
session.add(profile)
for key, value in data.items():
if key in allowed:
setattr(profile, key, value)
await session.commit()
await session.refresh(profile)
return profile
async def append_observations(user_id: int, bullets: str) -> None:
"""
Append a new dated observation entry from the day's briefing closeout.
Automatically triggers consolidation when the raw list grows large.
"""
if not bullets.strip():
return
async with async_session() as session:
result = await session.execute(
select(UserProfile).where(UserProfile.user_id == user_id)
)
profile = result.scalar_one_or_none()
if profile is None:
profile = UserProfile(user_id=user_id)
session.add(profile)
existing: list = list(profile.observations_raw or [])
existing.append({
"date": date.today().isoformat(),
"bullets": bullets.strip(),
})
# Keep at most 60 raw entries as a rolling window
profile.observations_raw = existing[-60:]
profile.observations_updated_at = datetime.now(timezone.utc)
await session.commit()
raw_count = len(profile.observations_raw or [])
logger.info("Appended observations for user %d (%d raw entries)", user_id, raw_count)
if raw_count >= _CONSOLIDATION_THRESHOLD:
asyncio.create_task(_consolidate_observations(user_id))
async def consolidate_observations(user_id: int) -> str:
"""Public entry point to manually trigger observation consolidation."""
return await _consolidate_observations(user_id)
async def clear_learned_data(user_id: int) -> None:
"""Reset all learned observations and summary for a user."""
async with async_session() as session:
result = await session.execute(
select(UserProfile).where(UserProfile.user_id == user_id)
)
profile = result.scalar_one_or_none()
if profile:
profile.learned_summary = None
profile.observations_raw = []
profile.observations_updated_at = None
await session.commit()
async def _consolidate_observations(user_id: int) -> str:
"""
LLM pass: synthesise all raw observation bullets into an updated
learned_summary paragraph. Prunes raw entries older than 30 days afterwards.
"""
from fabledassistant.config import Config
from fabledassistant.services.briefing_pipeline import _llm_synthesise
from fabledassistant.services.settings import get_setting
async with async_session() as session:
result = await session.execute(
select(UserProfile).where(UserProfile.user_id == user_id)
)
profile = result.scalar_one_or_none()
if not profile or not profile.observations_raw:
return ""
observations = list(profile.observations_raw)
existing_summary = profile.learned_summary or ""
obs_text = "\n\n".join(
f"[{entry['date']}]\n{entry['bullets']}"
for entry in observations
)
system = (
"You are synthesising preference observations into a concise user profile summary. "
"Consolidate the observations into 3-6 factual sentences describing the user's patterns, "
"preferences, and habits. Be specific and useful for a personal assistant. "
"Merge with any existing summary, removing duplicates and outdated information. "
"Output only the consolidated summary paragraph — no preamble, no bullet points."
)
user_prompt = ""
if existing_summary:
user_prompt += f"Existing summary:\n{existing_summary}\n\n"
user_prompt += f"New observations:\n{obs_text}"
model = await get_setting(user_id, "default_model", Config.OLLAMA_MODEL)
new_summary = await _llm_synthesise(system, user_prompt, model)
if new_summary:
cutoff = (date.today() - timedelta(days=30)).isoformat()
async with async_session() as session:
result = await session.execute(
select(UserProfile).where(UserProfile.user_id == user_id)
)
profile = result.scalar_one_or_none()
if profile:
profile.learned_summary = new_summary
profile.observations_raw = [
o for o in (profile.observations_raw or [])
if o.get("date", "") >= cutoff
]
await session.commit()
logger.info("Consolidated observations for user %d", user_id)
return new_summary
async def build_profile_context(user_id: int) -> str:
"""
Build a formatted context string from the user's structured profile
for injection into LLM system prompts (briefing and chat).
Returns an empty string if no meaningful data is set.
"""
profile = await get_profile(user_id)
parts: list[str] = []
if profile.display_name:
parts.append(f"User's name: {profile.display_name}")
if profile.job_title or profile.industry:
job = " in ".join(filter(None, [profile.job_title, profile.industry]))
parts.append(f"Occupation: {job}")
if profile.expertise_level and profile.expertise_level != "intermediate":
parts.append(
f"Expertise level: {profile.expertise_level} — calibrate explanation depth accordingly"
)
if profile.response_style or profile.tone:
style = profile.response_style or "balanced"
tone = profile.tone or "casual"
parts.append(f"Preferred response style: {style}, tone: {tone}")
if profile.interests:
parts.append(f"Interests: {', '.join(profile.interests)}")
if profile.work_schedule:
sched = profile.work_schedule
days = ", ".join(sched.get("days") or []) or "weekdays"
start = sched.get("start", "9:00")
end = sched.get("end", "17:00")
parts.append(f"Work schedule: {days}, {start}{end}")
if profile.learned_summary:
parts.append(f"What the assistant has learned about this user: {profile.learned_summary}")
return "\n".join(parts)