E3: add Graphiti retrieval branch to dream.py, E3 experiment script with blinding

scripts/dream.py

@@ -47,6 +47,7 @@ MODE_RANGES = {
     "late-rem":  (0.22, 0.42),
     "lucid":     (0.32, 0.72),
 }
+DREAMER_VERSION = "1.1"  # 1.0=original exclusion logic; 1.1=score-band exclusion
 
 # ─── Prompt versioning ──────────────────────────────────────────────────────
 # Bump the relevant constant manually when changing a prompt.
@@ -63,6 +64,11 @@ def prompt_hash(prompts: list[str]) -> str:
     combined = "".join(prompts)
     return hashlib.md5(combined.encode()).hexdigest()[:8]
 
+def extract_folder(source_path):
+    """Extract top-level Nextcloud folder from source path."""
+    parts = source_path.replace("\\", "/").split("/")
+    return parts[0] if parts else "unknown"
+
 # ─── Stage 1: Observe ───────────────────────────────────────────────────────
 
 def observe_corpus():
@@ -104,7 +110,55 @@ def get_recent_conversation_topics(days=14):
 
 # ─── Stage 2: Retrieve ──────────────────────────────────────────────────────
 
+
+def retrieve_graphiti(mode, task=None, n_results=8):
+    """E3 experiment — Graphiti substrate retrieval.
+    Queries Graphiti /search endpoint instead of pgvector.
+    Returns chunks in same format as retrieve() for pipeline compatibility.
+    Note: content is Graphiti facts (synthesized relationships), not raw chunks.
+    """
+    import requests as req_lib
+    if task:
+        query = task
+    elif mode == "late-rem":
+        delta = observe_corpus()
+        topics = delta.get("recent_topics", [])
+        query = topics[0] if topics else "practice place memory making"
+    elif mode == "early-rem":
+        query = "career decision personal change what matters next"
+    else:
+        query = "research fabrication teaching practice recent work"
+
+    try:
+        resp = req_lib.get(
+            "http://localhost:8001/search",
+            params={"query": query, "limit": n_results, "group_id": "aaron"},
+            timeout=30,
+        )
+        resp.raise_for_status()
+        results = resp.json().get("results", [])
+        chunks = []
+        for r in results:
+            fact = r.get("fact", "")
+            if not fact.strip():
+                continue
+            chunks.append({
+                "source": r.get("source", "graphiti"),
+                "content": fact,
+                "relevance": r.get("score", 0.5),
+                "similarity": r.get("score", 0.5),
+            })
+        return chunks
+    except Exception as e:
+        print(f"[Graphiti retrieval error: {e}] — falling back to empty.")
+        return []
+
 def retrieve(mode, task=None, n_results=8, excluded_sources=None):
+    # E3 experiment: DREAMER_SUBSTRATE=graphiti routes retrieval to Graphiti /search
+    # Default behavior: pgvector similarity search (unchanged)
+    substrate = os.getenv("DREAMER_SUBSTRATE", "pgvector")
+    if substrate == "graphiti":
+        return retrieve_graphiti(mode, task=task, n_results=n_results)
     from sentence_transformers import SentenceTransformer
     embedder = SentenceTransformer("all-MiniLM-L6-v2")
     low, high = MODE_RANGES[mode]
@@ -153,6 +207,7 @@ def retrieve(mode, task=None, n_results=8, excluded_sources=None):
                 "source": source or "unknown",
                 "content": doc,
                 "relevance": similarity,
+                "similarity": similarity,
             })
             seen_sources.add(source)
             if len(chunks) >= n_results:
@@ -361,6 +416,7 @@ def write_manifest(date_str, stage_data, corpus_data):
     manifest = {
         "date": date_str,
         "prompt_sig": prompt_signature(),
+        "dreamer_version": DREAMER_VERSION,
         "prompt_hash": prompt_hash([
             synthesize_nrem.__doc__ or "",
             synthesize_early_rem.__doc__ or "",
@@ -401,6 +457,8 @@ def dream_pipeline():
     print("\n[NREM] Retrieving...")
     nrem_chunks = retrieve("nrem", excluded_sources=previously_retrieved | session_retrieved)
     session_retrieved.update(c["source"] for c in nrem_chunks)
+    # Track sources that scored above Early REM ceiling — these are the only ones Early REM should exclude
+    nrem_high_sources = {c["source"] for c in nrem_chunks if c["similarity"] > 0.55}
     if not nrem_chunks:
         print("[NREM] No suitable chunks — aborting pipeline")
         return None
@@ -408,12 +466,17 @@ def dream_pipeline():
     print(f"[NREM] Retrieved {len(nrem_chunks)} chunks. Synthesizing...")
     nrem_output = synthesize_nrem(nrem_chunks)
     nrem_file = deliver(nrem_output, "nrem")
+    nrem_sources = [c["source"] for c in nrem_chunks]
+    nrem_folders = list({extract_folder(s) for s in nrem_sources})
     stage_data = {
         "nrem": {
             "chunks_retrieved": len(nrem_chunks),
             "avg_similarity": round(sum(c["relevance"] for c in nrem_chunks) / len(nrem_chunks), 3),
             "query": "research fabrication teaching practice recent work",
             "word_count": len(nrem_output.split()),
+            "sources": nrem_sources,
+            "distinct_folders": nrem_folders,
+            "folder_count": len(nrem_folders),
             "status": "ok",
         }
     }
@@ -421,7 +484,9 @@ def dream_pipeline():
 
     # ── Stage 2: Early REM — informed by NREM ──────────────────────────────
     print("\n[Early REM] Retrieving...")
-    early_chunks = retrieve("early-rem", excluded_sources=previously_retrieved | session_retrieved)
+    # Early REM excludes previously retrieved + NREM high-scorers only (not full session_retrieved)
+    # Sources that scored in Early REM band during NREM remain available
+    early_chunks = retrieve("early-rem", excluded_sources=previously_retrieved | nrem_high_sources)
     session_retrieved.update(c["source"] for c in early_chunks)
     if not early_chunks:
         print("[Early REM] No suitable chunks — skipping")
@@ -430,11 +495,16 @@ def dream_pipeline():
         print(f"[Early REM] Retrieved {len(early_chunks)} chunks. Synthesizing with NREM context...")
         early_rem_output = synthesize_early_rem(early_chunks, nrem_output)
         deliver(early_rem_output, "early-rem")
+        early_sources = [c["source"] for c in early_chunks]
+        early_folders = list({extract_folder(s) for s in early_sources})
         stage_data["early_rem"] = {
             "chunks_retrieved": len(early_chunks),
             "avg_similarity": round(sum(c["relevance"] for c in early_chunks) / len(early_chunks), 3),
             "query": "career decision personal change what matters next",
             "word_count": len(early_rem_output.split()),
+            "sources": early_sources,
+            "distinct_folders": early_folders,
+            "folder_count": len(early_folders),
             "status": "ok",
         }
         print(f"[Early REM] Done.\n{early_rem_output[:200]}...")
@@ -450,11 +520,22 @@ def dream_pipeline():
         print(f"[Late REM] Retrieved {len(late_chunks)} chunks. Synthesizing with full context...")
         late_rem_output = synthesize_late_rem(late_chunks, nrem_output, early_rem_output)
         deliver(late_rem_output, "late-rem")
+        late_sources = [c["source"] for c in late_chunks]
+        late_folders = [extract_folder(s) for s in late_sources]
+        cross_domain_pairs = sum(
+            1 for i in range(len(late_folders))
+            for j in range(i+1, len(late_folders))
+            if late_folders[i] != late_folders[j]
+        )
         stage_data["late_rem"] = {
             "chunks_retrieved": len(late_chunks),
             "avg_similarity": round(sum(c["relevance"] for c in late_chunks) / len(late_chunks), 3),
             "query": "practice place memory making",
             "word_count": len(late_rem_output.split()),
+            "sources": late_sources,
+            "distinct_folders": list(set(late_folders)),
+            "folder_count": len(set(late_folders)),
+            "cross_domain_pairs": cross_domain_pairs,
             "status": "ok",
         }
         print(f"[Late REM] Done.\n{late_rem_output[:200]}...")
@@ -474,10 +555,18 @@ def dream_pipeline():
     print(f"{'='*60}")
 
     # Write manifest
+    all_session_sources = list(session_retrieved)
+    all_session_folders = list({extract_folder(s) for s in all_session_sources})
     corpus_data = {
         "total_chunks": delta.get("new_chunks", 0),
         "new_chunks_since_last_dream": delta.get("new_chunks", 0),
         "days_since_last_dream": round(delta.get("days_since_dream", 0), 2),
+        "substrate": "pgvector",
+        "aggregate": {
+            "total_distinct_sources": len(all_session_sources),
+            "total_distinct_folders": len(all_session_folders),
+            "folders_touched": all_session_folders,
+        }
     }
     write_manifest(datetime.now().strftime("%Y-%m-%d"), stage_data, corpus_data)
 

scripts/experiments/e3_dreamer_substrate.py

@@ -0,0 +1,222 @@
+#!/usr/bin/env python3
+"""
+E3 — Dreamer Substrate Comparison Experiment
+Runs dream pipeline twice: once against pgvector (Condition A),
+once against Graphiti (Condition B). Implements blinding per protocol.
+
+Usage:
+    python3 scripts/experiments/e3_dreamer_substrate.py
+
+Outputs:
+    Journal/Dreams/e3-dream-X-{mode}.md  (one condition, identity concealed)
+    Journal/Dreams/e3-dream-Y-{mode}.md  (other condition, identity concealed)
+    Journal/Dreams/e3-blind-mapping.json (kept unread until ratings complete)
+    Journal/Dreams/e3-manifest.json      (quantitative metrics for both conditions)
+"""
+
+import os
+import sys
+import json
+import random
+import requests
+from pathlib import Path
+from datetime import datetime
+from dotenv import load_dotenv
+
+load_dotenv(Path.home() / "aaronai" / ".env", override=True)
+
+NEXTCLOUD_URL      = os.getenv("NEXTCLOUD_URL", "https://nextcloud.aaronnelson.studio")
+NEXTCLOUD_USER     = os.getenv("NEXTCLOUD_USER", "aaron")
+NEXTCLOUD_PASSWORD = os.getenv("NEXTCLOUD_PASSWORD", "")
+DREAMS_WEBDAV_BASE = f"{NEXTCLOUD_URL}/remote.php/dav/files/{NEXTCLOUD_USER}/Journal/Dreams"
+
+auth = (NEXTCLOUD_USER, NEXTCLOUD_PASSWORD)
+
+def webdav_put(url, content_str):
+    requests.put(url, data=content_str.encode("utf-8"), auth=auth, timeout=60)
+
+def webdav_mkcol(url):
+    requests.request("MKCOL", url, auth=auth, timeout=10)
+
+def run_condition(label, substrate, modes=("nrem", "early-rem", "late-rem", "synthesis")):
+    """Run the dream pipeline for one substrate condition.
+    Returns dict of {mode: dream_text} and metrics.
+    """
+    print(f"\n{'='*60}")
+    print(f"Running Condition {label} — substrate: {substrate}")
+    print(f"{'='*60}")
+
+    os.environ["DREAMER_SUBSTRATE"] = substrate
+
+    # Import dream module fresh for this run
+    if "dream" in sys.modules:
+        del sys.modules["dream"]
+    sys.path.insert(0, str(Path.home() / "aaronai" / "scripts"))
+    import dream as d
+
+    # Run pipeline stages manually to capture outputs
+    outputs = {}
+    metrics = {}
+
+    # Observe corpus delta
+    delta = d.observe_corpus()
+    print(f"Corpus delta: {delta.get('new_chunks', 0)} new chunks")
+
+    previously_retrieved = set()
+    session_retrieved = set()
+    nrem_high_sources = set()
+
+    # NREM
+    print("\n[NREM] Retrieving...")
+    nrem_chunks = d.retrieve("nrem", excluded_sources=previously_retrieved | session_retrieved)
+    session_retrieved.update(c["source"] for c in nrem_chunks)
+    nrem_high_sources = {c["source"] for c in nrem_chunks if c.get("similarity", 0) > 0.55}
+    if not nrem_chunks:
+        print("[NREM] No chunks — aborting.")
+        return None, None
+    nrem_output = d.synthesize_nrem(nrem_chunks)
+    outputs["nrem"] = nrem_output
+    metrics["nrem"] = {
+        "chunks_retrieved": len(nrem_chunks),
+        "sources": [c["source"] for c in nrem_chunks],
+        "word_count": len(nrem_output.split()),
+    }
+    print(f"[NREM] Done. {len(nrem_chunks)} chunks, {metrics['nrem']['word_count']} words.")
+
+    # Early REM
+    print("\n[Early REM] Retrieving...")
+    early_chunks = d.retrieve("early-rem", excluded_sources=previously_retrieved | nrem_high_sources)
+    session_retrieved.update(c["source"] for c in early_chunks)
+    if early_chunks:
+        early_output = d.synthesize_early_rem(early_chunks, nrem_output)
+        outputs["early-rem"] = early_output
+        metrics["early-rem"] = {
+            "chunks_retrieved": len(early_chunks),
+            "sources": [c["source"] for c in early_chunks],
+            "word_count": len(early_output.split()),
+        }
+        print(f"[Early REM] Done. {len(early_chunks)} chunks.")
+    else:
+        early_output = nrem_output
+        outputs["early-rem"] = None
+        metrics["early-rem"] = {"chunks_retrieved": 0, "sources": [], "word_count": 0}
+        print("[Early REM] No chunks — using NREM fallback.")
+
+    # Late REM
+    print("\n[Late REM] Retrieving...")
+    late_chunks = d.retrieve("late-rem", excluded_sources=previously_retrieved | session_retrieved)
+    session_retrieved.update(c["source"] for c in late_chunks)
+    if late_chunks:
+        late_output = d.synthesize_late_rem(late_chunks, nrem_output, early_output)
+        outputs["late-rem"] = late_output
+        late_sources = [c["source"] for c in late_chunks]
+        metrics["late-rem"] = {
+            "chunks_retrieved": len(late_chunks),
+            "sources": late_sources,
+            "word_count": len(late_output.split()),
+        }
+        print(f"[Late REM] Done. {len(late_chunks)} chunks.")
+    else:
+        late_output = early_output
+        outputs["late-rem"] = None
+        metrics["late-rem"] = {"chunks_retrieved": 0, "sources": [], "word_count": 0}
+        print("[Late REM] No chunks — using fallback.")
+
+    # Synthesis
+    print("\n[Synthesis] Integrating...")
+    synthesis_output = d.synthesize_final(nrem_output, early_output, late_output)
+    outputs["synthesis"] = synthesis_output
+    metrics["synthesis"] = {"word_count": len(synthesis_output.split())}
+    print(f"[Synthesis] Done. {metrics['synthesis']['word_count']} words.")
+
+    return outputs, metrics
+
+
+def main():
+    date_str = datetime.now().strftime("%Y-%m-%d")
+
+    # Randomize which condition gets label X vs Y for blinding
+    conditions = [("pgvector", "A"), ("graphiti", "B")]
+    labels = ["X", "Y"]
+    random.shuffle(labels)
+    assignment = {conditions[0][0]: labels[0], conditions[1][0]: labels[1]}
+
+    mapping = {
+        "date": date_str,
+        "X": assignment["pgvector"] == "X" and "pgvector" or "graphiti",
+        "Y": assignment["pgvector"] == "Y" and "pgvector" or "graphiti",
+        "reveal_after_rating": True,
+    }
+    # Simpler mapping
+    mapping = {
+        "date": date_str,
+        "X_is": "pgvector" if assignment["pgvector"] == "X" else "graphiti",
+        "Y_is": "pgvector" if assignment["pgvector"] == "Y" else "graphiti",
+        "reveal_after_rating": True,
+    }
+
+    print(f"E3 Dreamer Substrate Comparison — {date_str}")
+    print(f"Blind mapping written (do not read until ratings complete)")
+
+    # Write blind mapping to Nextcloud immediately
+    webdav_mkcol(DREAMS_WEBDAV_BASE)
+    mapping_url = f"{DREAMS_WEBDAV_BASE}/e3-blind-mapping.json"
+    webdav_put(mapping_url, json.dumps(mapping, indent=2))
+    print(f"Mapping written to Journal/Dreams/e3-blind-mapping.json")
+
+    all_metrics = {}
+
+    # Run both conditions
+    for substrate, condition_label in conditions:
+        blind_label = assignment[substrate]
+        outputs, metrics = run_condition(condition_label, substrate)
+
+        if outputs is None:
+            print(f"Condition {condition_label} failed — aborting E3.")
+            return
+
+        all_metrics[blind_label] = {
+            "substrate": "CONCEALED",  # concealed until reveal
+            "stages": metrics,
+        }
+
+        # Write dream files with blind labels
+        modes_map = {
+            "nrem": "nrem",
+            "early-rem": "early-rem",
+            "late-rem": "late-rem",
+            "synthesis": "synthesis",
+        }
+        for mode, text in outputs.items():
+            if text is None:
+                continue
+            filename = f"e3-dream-{blind_label}-{mode}.md"
+            header = f"# E3 Dream — {blind_label} — {mode.upper()} — {date_str}\n"
+            header += f"*Substrate concealed until rating complete*\n\n---\n\n"
+            file_content = header + text
+            url = f"{DREAMS_WEBDAV_BASE}/{filename}"
+            webdav_put(url, file_content)
+            print(f"Written: Journal/Dreams/{filename}")
+
+    # Write quantitative metrics (substrate concealed)
+    e3_manifest = {
+        "date": date_str,
+        "experiment": "E3",
+        "protocol": "E3-dreamer-substrate-comparison-protocol.md",
+        "blind_labels": ["X", "Y"],
+        "substrate_concealed": True,
+        "metrics": all_metrics,
+        "rating_dimensions": ["Texture", "Reach", "Specificity", "Forward-facing"],
+        "rating_scale": "1-5 per dimension per dream",
+    }
+    manifest_url = f"{DREAMS_WEBDAV_BASE}/e3-manifest.json"
+    webdav_put(manifest_url, json.dumps(e3_manifest, indent=2))
+    print(f"\nE3 manifest written to Journal/Dreams/e3-manifest.json")
+    print(f"\n{'='*60}")
+    print("E3 runs complete. Eight dream files written.")
+    print("Rate all 8 dreams before reading e3-blind-mapping.json")
+    print(f"{'='*60}")
+
+
+if __name__ == "__main__":
+    main()

scripts/ingest.py

@@ -143,6 +143,8 @@ def ingest_file(filepath):
         print(f"  Indexed {len(chunks)} chunks: {path.name}")
 
         # Enqueue for Stage 2 → Stage 3 (Graphiti pipeline)
+        # SKIP_STAGE2_ENQUEUE env var set by migration scripts to prevent bulk enqueue
+        if not os.getenv("SKIP_STAGE2_ENQUEUE"):
             enqueue_stage2(path.name, text)
 
         return len(chunks)

scripts/stage2_worker.py

@@ -21,7 +21,6 @@ logging.basicConfig(
     level=logging.INFO,
     format="%(asctime)s [stage2] %(levelname)s %(message)s",
     handlers=[
-        logging.StreamHandler(),
         logging.FileHandler("/var/log/aaronai/stage2.log", mode="a"),
     ]
 )

scripts/stage3_worker.py

@@ -24,7 +24,6 @@ logging.basicConfig(
     level=logging.INFO,
     format="%(asctime)s [stage3] %(levelname)s %(message)s",
     handlers=[
-        logging.StreamHandler(),
         logging.FileHandler("/var/log/aaronai/stage3.log", mode="a"),
     ]
 )

scripts/watcher.py

@@ -1,62 +1,214 @@
+"""
+Aaron AI Watcher — Stage 1 of the encoding pipeline.
+
+Watches the Nextcloud directory for new or changed files.
+On detection, chunks + embeds documents in-process (no subprocess),
+then enqueues to stage_2_queue for async cascade processing.
+
+Design principles:
+- Embedding model loaded ONCE at startup, reused across all ingest runs
+- In-process ingest (no subprocess) — eliminates per-run model reload memory spike
+- Missed-file recovery on startup — ingests anything new since last state
+- Heartbeat file updated every loop tick — enables external health monitoring
+- Parity principle: no filtering, no decisions, faithful capture
+- Does NOT enqueue to stage_2_queue during bulk migration (SKIP_STAGE2_ENQUEUE env var)
+
+Architecture: Stage 1 (watcher) -> stage_2_queue -> Stage 2 (Mistral) -> stage_3_queue -> Stage 3 (Graphiti)
+"""
+
+import os
 import time
-import subprocess
-import logging
 import json
+import hashlib
+import logging
 import threading
 from pathlib import Path
+
+import psycopg2
+from dotenv import load_dotenv
+from sentence_transformers import SentenceTransformer
 from watchdog.observers import Observer
 from watchdog.events import FileSystemEventHandler
 
+from docx import Document as DocxDocument
+from pypdf import PdfReader
+from pptx import Presentation
+
+load_dotenv(Path.home() / "aaronai" / ".env", override=True)
+
 NEXTCLOUD_PATH = "/home/aaron/nextcloud/data/data/aaron/files"
-INGEST_SCRIPT = "/home/aaron/aaronai/scripts/ingest.py"
-PYTHON = "/home/aaron/aaronai/venv/bin/python3"
 LOG_FILE       = "/home/aaron/aaronai/watcher.log"
 STATE_FILE     = "/home/aaron/aaronai/watcher_state.json"
-
-SUPPORTED = {'.pdf', '.docx', '.pptx', '.txt', '.md'}
-DEBOUNCE_SECONDS = 120
 STATUS_FILE    = "/home/aaron/aaronai/watcher_status.json"
+HEARTBEAT_FILE = "/home/aaron/aaronai/watcher_heartbeat"
+
+SUPPORTED        = {".pdf", ".docx", ".pptx", ".txt", ".md"}
+DEBOUNCE_SECONDS = 120
+CHUNK_SIZE       = 500
+CHUNK_OVERLAP    = 50
+EMBED_MODEL      = "all-MiniLM-L6-v2"
+
+PG_DSN = os.getenv("PG_DSN")
 
 logging.basicConfig(
     level=logging.INFO,
-    format='%(asctime)s - %(message)s',
-    handlers=[
-        logging.FileHandler(LOG_FILE),
-        logging.StreamHandler()
-    ]
+    format="%(asctime)s [watcher] %(levelname)s %(message)s",
+    handlers=[logging.FileHandler(LOG_FILE)],
 )
+log = logging.getLogger("watcher")
 
-ingestion_state = {
-    "status": "idle",
-    "message": "",
-    "file_count": 0,
-    "started_at": None,
-    "finished_at": None,
-    "last_error": "",
-}
 ingestion_lock  = threading.Lock()
+ingestion_state = {
+    "status": "idle", "message": "", "file_count": 0,
+    "started_at": None, "finished_at": None, "last_error": "",
+}
 ingestion_thread = None
 
 
-def set_ingestion_state(**kwargs):
-    with ingestion_lock:
-        ingestion_state.update(kwargs)
+def load_embedder():
+    log.info(f"Loading embedding model: {EMBED_MODEL}")
+    model = SentenceTransformer(EMBED_MODEL)
+    log.info("Embedding model ready.")
+    return model
 
 
-def load_state():
+def get_pg():
+    return psycopg2.connect(PG_DSN)
+
+
+def extract_text(path: Path) -> str:
+    suffix = path.suffix.lower()
+    try:
+        if suffix == ".docx":
+            doc = DocxDocument(path)
+            return "\n".join(p.text for p in doc.paragraphs if p.text.strip())
+        elif suffix == ".pdf":
+            reader = PdfReader(path)
+            return "".join(
+                page.extract_text() + "\n"
+                for page in reader.pages if page.extract_text()
+            )
+        elif suffix == ".pptx":
+            prs = Presentation(path)
+            return "\n".join(
+                shape.text for slide in prs.slides
+                for shape in slide.shapes
+                if hasattr(shape, "text") and shape.text.strip()
+            )
+        elif suffix in {".txt", ".md"}:
+            return path.read_text(encoding="utf-8", errors="ignore")
+    except Exception as e:
+        log.warning(f"Text extraction failed for {path.name}: {e}")
+    return ""
+
+
+def chunk_text(text: str) -> list:
+    words = text.split()
+    chunks = []
+    start = 0
+    while start < len(words):
+        chunk = " ".join(words[start:start + CHUNK_SIZE])
+        if chunk.strip():
+            chunks.append(chunk)
+        start += CHUNK_SIZE - CHUNK_OVERLAP
+    return chunks
+
+
+def make_chunk_id(filepath: Path, chunk_index: int) -> str:
+    return hashlib.md5(str(filepath).encode()).hexdigest()[:8] + f"_{chunk_index}"
+
+
+def enqueue_stage2(source: str, full_text: str):
+    if os.getenv("SKIP_STAGE2_ENQUEUE"):
+        return
+    try:
+        pg = get_pg()
+        cur = pg.cursor()
+        cur.execute("""
+            INSERT INTO stage_2_queue (source, full_text, char_length)
+            VALUES (%s, %s, %s)
+            ON CONFLICT (source) DO UPDATE SET
+                full_text    = EXCLUDED.full_text,
+                char_length  = EXCLUDED.char_length,
+                enqueued_at  = NOW(),
+                completed_at = NULL,
+                failed_at    = NULL,
+                attempts     = 0
+        """, (source, full_text[:50000], len(full_text)))
+        pg.commit()
+        pg.close()
+    except Exception as e:
+        log.warning(f"Stage 2 enqueue failed (non-fatal): {e}")
+
+
+def ingest_file(filepath: Path, embedder) -> int:
+    if filepath.name.startswith(("~$", ".")):
+        return 0
+    if filepath.suffix.lower() not in SUPPORTED:
+        return 0
+    text = extract_text(filepath)
+    if not text.strip():
+        return 0
+    chunks = chunk_text(text)
+    if not chunks:
+        return 0
+    try:
+        embeddings = embedder.encode(chunks).tolist()
+    except Exception as e:
+        log.error(f"Embedding failed for {filepath.name}: {e}")
+        return 0
+    source = filepath.name
+    try:
+        pg  = get_pg()
+        cur = pg.cursor()
+        for i, (chunk, embedding) in enumerate(zip(chunks, embeddings)):
+            chunk_id = make_chunk_id(filepath, i)
+            cur.execute("""
+                INSERT INTO embeddings (id, document, embedding, source, type, created_at, metadata)
+                VALUES (%s, %s, %s::vector, %s, %s, NOW(), %s)
+                ON CONFLICT (id) DO UPDATE SET
+                    document  = EXCLUDED.document,
+                    embedding = EXCLUDED.embedding,
+                    source    = EXCLUDED.source,
+                    metadata  = EXCLUDED.metadata
+            """, (chunk_id, chunk, embedding, source, "document",
+                  json.dumps({"source": source, "filepath": str(filepath)})))
+        pg.commit()
+        pg.close()
+    except Exception as e:
+        log.error(f"pgvector write failed for {filepath.name}: {e}")
+        return 0
+    log.info(f"Indexed {len(chunks)} chunks: {filepath.name}")
+    enqueue_stage2(source, text)
+    return len(chunks)
+
+
+def ingest_files(paths: list, embedder, state: dict) -> dict:
+    total = 0
+    for path in paths:
+        count = ingest_file(path, embedder)
+        total += count
+        state[str(path)] = str(path.stat().st_mtime)
+    log.info(f"Ingestion complete. {total} chunks across {len(paths)} files.")
+    return state
+
+
+def load_state() -> dict:
     if Path(STATE_FILE).exists():
+        try:
             with open(STATE_FILE) as f:
                 return json.load(f)
+        except Exception:
+            pass
     return {}
 
 
-def save_state(state):
-    with open(STATE_FILE, 'w') as f:
+def save_state(state: dict):
+    with open(STATE_FILE, "w") as f:
         json.dump(state, f)
 
 
-def get_changed_files():
-    state = load_state()
+def get_changed_files(state: dict) -> list:
     changed = []
     root = Path(NEXTCLOUD_PATH)
     for path in root.rglob("*"):
@@ -64,84 +216,81 @@ def get_changed_files():
             continue
         if path.suffix.lower() not in SUPPORTED:
             continue
-        if path.name.startswith('.') or path.name.startswith('~$'):
+        if path.name.startswith((".", "~$")):
             continue
-        mtime = str(path.stat().st_mtime)
-        key = str(path)
-        if state.get(key) != mtime:
+        if "Admin/Backups" in str(path) or "Backups" in path.parts:
+            continue
+        if "Journal/Media" in str(path):
+            continue
+        if state.get(str(path)) != str(path.stat().st_mtime):
             changed.append(path)
-    return changed, state
+    return changed
 
 
-def run_ingestion():
-    changed, state = get_changed_files()
+def set_ingestion_state(**kwargs):
+    with ingestion_lock:
+        ingestion_state.update(kwargs)
+
+
+def write_status(handler):
+    with ingestion_lock:
+        status = {
+            "running": True, "timestamp": time.time(),
+            "pending": handler.pending, "last_event": handler.last_event,
+            "ingestion": dict(ingestion_state),
+        }
+    try:
+        with open(STATUS_FILE, "w") as f:
+            json.dump(status, f)
+    except Exception:
+        pass
+
+
+def write_heartbeat():
+    try:
+        Path(HEARTBEAT_FILE).write_text(str(time.time()))
+    except Exception:
+        pass
+
+
+def run_ingestion(embedder):
+    state   = load_state()
+    changed = get_changed_files(state)
     if not changed:
-        logging.info("No new or changed files detected — skipping ingestion.")
+        log.info("No new or changed files — skipping ingestion.")
         set_ingestion_state(status="idle", message="No changes detected", file_count=0)
         return
-
     count = len(changed)
-    logging.info(f"Found {count} new or changed files — starting ingestion...")
+    log.info(f"Found {count} new or changed files — starting ingestion...")
     set_ingestion_state(
-        status="ingesting",
-        message=f"Ingesting {count} file(s)...",
-        file_count=count,
-        started_at=time.time(),
-        finished_at=None,
-        last_error="",
+        status="ingesting", message=f"Ingesting {count} file(s)...",
+        file_count=count, started_at=time.time(), finished_at=None, last_error="",
     )
-
     try:
-        result = subprocess.run(
-            [PYTHON, INGEST_SCRIPT, NEXTCLOUD_PATH],
-            capture_output=True,
-            text=True,
-            timeout=1800
-        )
-        if result.returncode == 0:
-            root = Path(NEXTCLOUD_PATH)
-            for path in root.rglob("*"):
-                if path.is_file() and path.suffix.lower() in SUPPORTED:
-                    state[str(path)] = str(path.stat().st_mtime)
+        state = ingest_files(changed, embedder, state)
         save_state(state)
-            logging.info("Ingestion complete. State updated.")
         set_ingestion_state(
             status="idle",
             message=f"Last run: ingested {count} file(s) successfully",
             finished_at=time.time(),
         )
-        else:
-            logging.error(f"Ingestion error: {result.stderr}")
-            set_ingestion_state(
-                status="error",
-                message="Ingestion failed — see log",
-                last_error=result.stderr[-300:],
-                finished_at=time.time(),
-            )
-    except subprocess.TimeoutExpired:
-        logging.error("Ingestion timed out.")
-        set_ingestion_state(
-            status="error",
-            message="Ingestion timed out (>30 min)",
-            last_error="TimeoutExpired",
-            finished_at=time.time(),
-        )
     except Exception as e:
-        logging.error(f"Ingestion failed: {e}")
+        log.error(f"Ingestion failed: {e}")
         set_ingestion_state(
-            status="error",
-            message=f"Ingestion exception: {e}",
-            last_error=str(e),
-            finished_at=time.time(),
+            status="error", message=f"Ingestion exception: {e}",
+            last_error=str(e), finished_at=time.time(),
         )
 
 
-def start_ingestion_thread():
+def start_ingestion_thread(embedder):
     global ingestion_thread
+    with ingestion_lock:
         if ingestion_thread and ingestion_thread.is_alive():
-        logging.info("Ingestion already running — skipping.")
+            log.info("Ingestion already running — skipping.")
             return
-    ingestion_thread = threading.Thread(target=run_ingestion, daemon=True)
+        ingestion_thread = threading.Thread(
+            target=run_ingestion, args=(embedder,), daemon=True
+        )
         ingestion_thread.start()
 
 
@@ -156,54 +305,72 @@ class IngestHandler(FileSystemEventHandler):
         path = Path(event.src_path)
         if path.suffix.lower() not in SUPPORTED:
             return
-        if path.name.startswith('.') or path.name.startswith('~$'):
+        if path.name.startswith((".", "~$")):
             return
-        if 'Admin/Backups' in str(path) or 'Backups' in path.parts:
+        if "Admin/Backups" in str(path) or "Backups" in path.parts:
             return
-        if 'Journal/Media' in str(path):
+        if "Journal/Media" in str(path):
             return
-        if event.event_type not in ('modified', 'created', 'moved'):
+        if event.event_type not in ("modified", "created", "moved"):
             return
-        logging.info(f"Event: {event.event_type} {event.src_path}")
+        log.info(f"Event: {event.event_type} {event.src_path}")
         self.pending    = True
         self.last_event = time.time()
 
 
-def write_status(handler):
-    with ingestion_lock:
-        status = {
-            "running": True,
-            "timestamp": time.time(),
-            "pending": handler.pending,
-            "last_event": handler.last_event,
-            "ingestion": dict(ingestion_state),
-        }
-    with open(STATUS_FILE, 'w') as f:
-        json.dump(status, f)
-
-
 def main():
-    logging.info("Aaron AI Watcher starting...")
-    logging.info(f"Watching: {NEXTCLOUD_PATH}")
+    log.info("Aaron AI Watcher starting...")
+    log.info(f"Watching: {NEXTCLOUD_PATH}")
+
+    embedder = load_embedder()
+
+    log.info("Startup scan: checking for files missed since last run...")
+    state  = load_state()
+    missed = get_changed_files(state)
+    if missed:
+        log.info(f"Startup recovery: {len(missed)} missed file(s) — ingesting now.")
+        set_ingestion_state(
+            status="ingesting",
+            message=f"Startup recovery: ingesting {len(missed)} missed file(s)...",
+            file_count=len(missed), started_at=time.time(),
+        )
+        try:
+            state = ingest_files(missed, embedder, state)
+            save_state(state)
+            set_ingestion_state(
+                status="idle",
+                message=f"Startup recovery complete: {len(missed)} file(s) ingested.",
+                finished_at=time.time(),
+            )
+        except Exception as e:
+            log.error(f"Startup recovery failed: {e}")
+            set_ingestion_state(status="error", message=str(e),
+                                last_error=str(e), finished_at=time.time())
+    else:
+        log.info("Startup scan: no missed files.")
 
     handler  = IngestHandler()
     observer = Observer()
     observer.schedule(handler, NEXTCLOUD_PATH, recursive=True)
     observer.start()
+    log.info("Observer started.")
 
     try:
         while True:
+            write_heartbeat()
             write_status(handler)
             if handler.pending:
                 elapsed = time.time() - handler.last_event
                 if elapsed >= DEBOUNCE_SECONDS:
                     handler.pending = False
-                    start_ingestion_thread()
+                    start_ingestion_thread(embedder)
             time.sleep(5)
     except KeyboardInterrupt:
+        log.info("KeyboardInterrupt — stopping.")
         observer.stop()
+
     observer.join()
-    logging.info("Watcher stopped.")
+    log.info("Watcher stopped.")
 
 
 if __name__ == "__main__":