SpeedyFoxAi/vera-ai-v2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

v2.0.1: Monthly curation now in curator_prompt.md, remove full_run_time/full_run_day config

Browse Source
This commit is contained in:
Vera-AI
2026-03-26 21:26:02 -05:00
parent f6affc9e01
commit 50874eeae9
7 changed files with 49 additions and 103 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
DOCKERHUB.md
View File

@@ -148,10 +148,8 @@ semantic_score_threshold = 0.6
run_time = "02:00" run_time = "02:00"
# Time for monthly full curation (HH:MM format) # Time for monthly full curation (HH:MM format)
full_run_time = "03:00"
# Day of month for full curation (1-28) # Day of month for full curation (1-28)
full_run_day = 1
# Model to use for curation # Model to use for curation
curator_model = "gpt-oss:120b" curator_model = "gpt-oss:120b"
@@ -308,7 +306,7 @@ docker run -d --name VeraAI -p 8080:11434 ...
| Feature | Description | | Feature | Description |
|---------|-------------| |---------|-------------|
| 🧠 **Persistent Memory** | Conversations stored in Qdrant, retrieved contextually | | 🧠 **Persistent Memory** | Conversations stored in Qdrant, retrieved contextually |
| 📅 **Monthly Curation** | Daily + monthly cleanup of raw memories | | 📅 **Monthly Curation** | Daily cleanup, auto-monthly on day 01 |
| 🔍 **4-Layer Context** | System + semantic + recent + current messages | | 🔍 **4-Layer Context** | System + semantic + recent + current messages |
| 👤 **Configurable UID/GID** | Match container user to host for permissions | | 👤 **Configurable UID/GID** | Match container user to host for permissions |
| 🌍 **Timezone Support** | Scheduler runs in your local timezone | | 🌍 **Timezone Support** | Scheduler runs in your local timezone |

7
README.md
View File

@@ -58,7 +58,7 @@ Every conversation is stored in Qdrant vector database and retrieved contextuall
| Feature | Description | | Feature | Description |
|---------|-------------| |---------|-------------|
| **🧠 Persistent Memory** | Conversations stored in Qdrant, retrieved contextually | | **🧠 Persistent Memory** | Conversations stored in Qdrant, retrieved contextually |
| **📅 Monthly Curation** | Daily + monthly cleanup of raw memories | | **📅 Smart Curation** | Daily cleanup, auto-monthly on day 01 |
| **🔍 4-Layer Context** | System + semantic + recent + current messages | | **🔍 4-Layer Context** | System + semantic + recent + current messages |
| **👤 Configurable UID/GID** | Match container user to host for permissions | | **👤 Configurable UID/GID** | Match container user to host for permissions |
| **🌍 Timezone Support** | Scheduler runs in your local timezone | | **🌍 Timezone Support** | Scheduler runs in your local timezone |
@@ -314,10 +314,8 @@ run_time = "02:00"
# Time for monthly full curation (HH:MM format, 24-hour) # Time for monthly full curation (HH:MM format, 24-hour)
# Processes ALL raw memories # Processes ALL raw memories
full_run_time = "03:00"
# Day of month for full curation (1-28) # Day of month for full curation (1-28)
full_run_day = 1
# Model to use for curation # Model to use for curation
# Should be a capable model for summarization # Should be a capable model for summarization
@@ -540,7 +538,8 @@ TZ=Europe/London # GMT/BST
curl -X POST http://localhost:11434/curator/run curl -X POST http://localhost:11434/curator/run
# Full curation (all raw memories) # Full curation (all raw memories)
curl -X POST "http://localhost:11434/curator/run?full=true" # Monthly mode is automatic on day 01
# curl -X POST http://localhost:11434/curator/run
``` ```
--- ---

5
app/config.py
View File

@@ -48,8 +48,7 @@ class Config:
semantic_search_turns: int = 2 semantic_search_turns: int = 2
semantic_score_threshold: float = 0.6 # Score threshold for semantic search semantic_score_threshold: float = 0.6 # Score threshold for semantic search
run_time: str = "02:00" # Daily curator time run_time: str = "02:00" # Daily curator time
full_run_time: str = "03:00" # Monthly full curator time # Monthly mode is detected by curator_prompt.md (day 01)
full_run_day: int = 1 # Day of month for full run (1st)
curator_model: str = "gpt-oss:120b" curator_model: str = "gpt-oss:120b"
debug: bool = False debug: bool = False
cloud: CloudConfig = field(default_factory=CloudConfig) cloud: CloudConfig = field(default_factory=CloudConfig)
@@ -103,8 +102,6 @@ class Config:
if "curator" in data: if "curator" in data:
config.run_time = data["curator"].get("run_time", config.run_time) config.run_time = data["curator"].get("run_time", config.run_time)
config.full_run_time = data["curator"].get("full_run_time", config.full_run_time)
config.full_run_day = data["curator"].get("full_run_day", config.full_run_day)
config.curator_model = data["curator"].get("curator_model", config.curator_model) config.curator_model = data["curator"].get("curator_model", config.curator_model)
if "cloud" in data: if "cloud" in data:

51
app/curator.py
View File

@@ -1,7 +1,8 @@
"""Memory curator - runs daily (recent 24h) and monthly (full DB) to clean and maintain memory database. """Memory curator - runs daily to clean and maintain memory database.
Creates INDIVIDUAL cleaned turns (one per raw turn), not merged summaries. On day 01 of each month, processes ALL raw memories (monthly mode).
Parses JSON response from curator_prompt.md format. Otherwise, processes recent 24h of raw memories (daily mode).
The prompt determines behavior based on current date.
""" """
import logging import logging
import os import os
@@ -23,7 +24,6 @@ STATIC_DIR = Path(os.environ.get("VERA_STATIC_DIR", "/app/static"))
def load_curator_prompt() -> str: def load_curator_prompt() -> str:
"""Load curator prompt from prompts directory.""" """Load curator prompt from prompts directory."""
# Try prompts directory first, then static for backward compatibility
prompts_path = PROMPTS_DIR / "curator_prompt.md" prompts_path = PROMPTS_DIR / "curator_prompt.md"
static_path = STATIC_DIR / "curator_prompt.md" static_path = STATIC_DIR / "curator_prompt.md"
@@ -42,16 +42,20 @@ class Curator:
self.ollama_host = ollama_host self.ollama_host = ollama_host
self.curator_prompt = load_curator_prompt() self.curator_prompt = load_curator_prompt()
async def run(self, full: bool = False): async def run(self):
"""Run the curation process. """Run the curation process.
Args: Automatically detects day 01 for monthly mode (processes ALL raw memories).
full: If True, process ALL raw memories (monthly full run). Otherwise runs daily mode (processes recent 24h only).
If False, process only recent 24h (daily run). The prompt determines behavior based on current date.
""" """
logger.info(f"Starting memory curation (full={full})...") current_date = datetime.utcnow()
is_monthly = current_date.day == 1
mode = "MONTHLY" if is_monthly else "DAILY"
logger.info(f"Starting memory curation ({mode} mode)...")
try: try:
current_date = datetime.utcnow().strftime("%Y-%m-%d") current_date_str = current_date.strftime("%Y-%m-%d")
# Get all memories (async) # Get all memories (async)
points, _ = await self.qdrant.client.scroll( points, _ = await self.qdrant.client.scroll(
@@ -77,15 +81,15 @@ class Curator:
logger.info(f"Found {len(raw_memories)} raw, {len(curated_memories)} curated") logger.info(f"Found {len(raw_memories)} raw, {len(curated_memories)} curated")
# Filter by time for daily runs, process all for full runs # Filter by time for daily mode, process all for monthly mode
if full: if is_monthly:
# Monthly full run: process ALL raw memories # Monthly full run: process ALL raw memories
recent_raw = raw_memories recent_raw = raw_memories
logger.info(f"FULL RUN: Processing all {len(recent_raw)} raw memories") logger.info(f"MONTHLY MODE: Processing all {len(recent_raw)} raw memories")
else: else:
# Daily run: process only recent 24h # Daily run: process only recent 24h
recent_raw = [m for m in raw_memories if self._is_recent(m, hours=24)] recent_raw = [m for m in raw_memories if self._is_recent(m, hours=24)]
logger.info(f"DAILY RUN: Processing {len(recent_raw)} recent raw memories") logger.info(f"DAILY MODE: Processing {len(recent_raw)} recent raw memories")
existing_sample = curated_memories[-50:] if len(curated_memories) > 50 else curated_memories existing_sample = curated_memories[-50:] if len(curated_memories) > 50 else curated_memories
@@ -96,10 +100,10 @@ class Curator:
raw_turns_text = self._format_raw_turns(recent_raw) raw_turns_text = self._format_raw_turns(recent_raw)
existing_text = self._format_existing_memories(existing_sample) existing_text = self._format_existing_memories(existing_sample)
prompt = self.curator_prompt.replace("{CURRENT_DATE}", current_date) prompt = self.curator_prompt.replace("{CURRENT_DATE}", current_date_str)
full_prompt = f"""{prompt} full_prompt = f"""{prompt}
## {'All' if full else 'Recent'} Raw Turns ({'full database' if full else 'last 24 hours'}): ## {'All' if is_monthly else 'Recent'} Raw Turns ({'full database' if is_monthly else 'last 24 hours'}):
{raw_turns_text} {raw_turns_text}
## Existing Memories (sample): ## Existing Memories (sample):
@@ -152,20 +156,12 @@ Remember: Respond with ONLY valid JSON. No markdown, no explanations, just the J
await self.qdrant.delete_points(raw_ids_to_delete) await self.qdrant.delete_points(raw_ids_to_delete)
logger.info(f"Deleted {len(raw_ids_to_delete)} processed raw memories") logger.info(f"Deleted {len(raw_ids_to_delete)} processed raw memories")
logger.info(f"Memory curation completed successfully (full={full})") logger.info(f"Memory curation completed successfully ({mode} mode)")
except Exception as e: except Exception as e:
logger.error(f"Error during curation: {e}") logger.error(f"Error during curation: {e}")
raise raise
async def run_full(self):
"""Run full curation (all raw memories). Convenience method."""
await self.run(full=True)
async def run_daily(self):
"""Run daily curation (recent 24h only). Convenience method."""
await self.run(full=False)
def _is_recent(self, memory: Dict, hours: int = 24) -> bool: def _is_recent(self, memory: Dict, hours: int = 24) -> bool:
"""Check if memory is within the specified hours.""" """Check if memory is within the specified hours."""
timestamp = memory.get("timestamp", "") timestamp = memory.get("timestamp", "")
@@ -236,7 +232,9 @@ Remember: Respond with ONLY valid JSON. No markdown, no explanations, just the J
except json.JSONDecodeError: except json.JSONDecodeError:
pass pass
json_match = re.search(r'```(?:json)?\s*([\s\S]*?)```', response) # Try to find JSON in code blocks
pattern = r'```(?:json)?\s*([\s\S]*?)```'
json_match = re.search(pattern, response)
if json_match: if json_match:
try: try:
return json.loads(json_match.group(1).strip()) return json.loads(json_match.group(1).strip())
@@ -248,7 +246,6 @@ Remember: Respond with ONLY valid JSON. No markdown, no explanations, just the J
async def _append_rule_to_file(self, filename: str, rule: str): async def _append_rule_to_file(self, filename: str, rule: str):
"""Append a permanent rule to a prompts file.""" """Append a permanent rule to a prompts file."""
# Try prompts directory first, then static for backward compatibility
prompts_path = PROMPTS_DIR / filename prompts_path = PROMPTS_DIR / filename
static_path = STATIC_DIR / filename static_path = STATIC_DIR / filename

54
app/main.py
View File

@@ -20,25 +20,19 @@ curator = None
async def run_curator(): async def run_curator():
"""Scheduled daily curator job (recent 24h).""" """Scheduled daily curator job.
global curator
logger.info("Starting daily memory curation...")
try:
await curator.run_daily()
logger.info("Daily memory curation completed successfully")
except Exception as e:
logger.error(f"Daily memory curation failed: {e}")
Runs every day at configured time. The curator itself detects
async def run_curator_full(): if it's day 01 (monthly mode) and processes all memories.
"""Scheduled monthly curator job (full database).""" Otherwise processes recent 24h only.
"""
global curator global curator
logger.info("Starting monthly full memory curation...") logger.info("Starting memory curation...")
try: try:
await curator.run_full() await curator.run()
logger.info("Monthly full memory curation completed successfully") logger.info("Memory curation completed successfully")
except Exception as e: except Exception as e:
logger.error(f"Monthly full memory curation failed: {e}") logger.error(f"Memory curation failed: {e}")
@asynccontextmanager @asynccontextmanager
@@ -59,23 +53,12 @@ async def lifespan(app: FastAPI):
ollama_host=config.ollama_host ollama_host=config.ollama_host
) )
# Schedule daily curator (recent 24h) # Schedule daily curator
# Note: Monthly mode is detected automatically by curator_prompt.md (day 01)
hour, minute = map(int, config.run_time.split(":")) hour, minute = map(int, config.run_time.split(":"))
scheduler.add_job(run_curator, "cron", hour=hour, minute=minute, id="daily_curator") scheduler.add_job(run_curator, "cron", hour=hour, minute=minute, id="daily_curator")
logger.info(f"Daily curator scheduled at {config.run_time}") logger.info(f"Daily curator scheduled at {config.run_time}")
# Schedule monthly full curator (all raw memories)
full_hour, full_minute = map(int, config.full_run_time.split(":"))
scheduler.add_job(
run_curator_full,
"cron",
day=config.full_run_day,
hour=full_hour,
minute=full_minute,
id="monthly_curator"
)
logger.info(f"Monthly full curator scheduled on day {config.full_run_day} at {config.full_run_time}")
scheduler.start() scheduler.start()
yield yield
@@ -141,16 +124,11 @@ async def proxy_all(request: Request, path: str):
@app.post("/curator/run") @app.post("/curator/run")
async def trigger_curator(full: bool = False): async def trigger_curator():
"""Manually trigger curator. """Manually trigger curator.
Args: The curator will automatically detect if it's day 01 (monthly mode)
full: If True, run full curation (all raw memories). and process all memories. Otherwise processes recent 24h.
If False (default), run daily curation (recent 24h).
""" """
if full: await run_curator()
await run_curator_full() return {"status": "curation completed"}
return {"status": "full curation completed"}
else:
await run_curator()
return {"status": "daily curation completed"}

21
config.toml
View File

@@ -1,21 +0,0 @@
[general]
ollama_host = "http://10.0.0.10:11434"
qdrant_host = "http://10.0.0.22:6333"
qdrant_collection = "memories"
embedding_model = "snowflake-arctic-embed2"
debug = false
[layers]
# Note: system_token_budget removed - system prompt is never truncated
semantic_token_budget = 25000
context_token_budget = 22000
semantic_search_turns = 2
semantic_score_threshold = 0.6
[curator]
# Daily curation: processes recent 24h of raw memories
run_time = "02:00"
# Monthly full curation: processes ALL raw memories
full_run_time = "03:00"
full_run_day = 1 # Day of month (1st)
curator_model = "gpt-oss:120b"

4
config/config.toml
View File

@@ -14,8 +14,6 @@ semantic_score_threshold = 0.6
[curator] [curator]
# Daily curation: processes recent 24h of raw memories # Daily curation: processes recent 24h of raw memories
# Monthly mode is detected automatically by curator_prompt.md (day 01)
run_time = "02:00" run_time = "02:00"
# Monthly full curation: processes ALL raw memories
full_run_time = "03:00"
full_run_day = 1 # Day of month (1st)
curator_model = "gpt-oss:120b" curator_model = "gpt-oss:120b"