aethergraph 0.1.0a3__py3-none-any.whl → 0.1.0a4__py3-none-any.whl

aethergraph/storage/vector_index/sqlite_index.py

@@ -3,13 +3,21 @@ from __future__ import annotations
 import asyncio
 import json
 from pathlib import Path
+import pickle
 import sqlite3
+import threading
 from typing import Any
 
 import numpy as np
 
 from aethergraph.contracts.storage.vector_index import VectorIndex
 
+try:
+    import faiss  # type: ignore
+except Exception:
+    faiss = None
+
+
 SCHEMA = """
 CREATE TABLE IF NOT EXISTS chunks (
     corpus_id TEXT,
@@ -17,45 +25,234 @@ CREATE TABLE IF NOT EXISTS chunks (
     meta_json TEXT,
     PRIMARY KEY (corpus_id, chunk_id)
 );
+
 CREATE TABLE IF NOT EXISTS embeddings (
-    corpus_id TEXT,
-    chunk_id  TEXT,
-    vec       BLOB,    -- np.float32 array bytes
-    norm      REAL,
+    corpus_id     TEXT,
+    chunk_id      TEXT,
+    vec           BLOB,    -- np.float32 array bytes
+    norm          REAL,
+    -- promoted / hot fields
+    scope_id      TEXT,
+    user_id       TEXT,
+    org_id        TEXT,
+    client_id     TEXT,
+    session_id    TEXT,
+    run_id        TEXT,
+    graph_id      TEXT,
+    node_id       TEXT,
+    kind          TEXT,
+    source        TEXT,
+    created_at_ts REAL,
     PRIMARY KEY (corpus_id, chunk_id)
 );
+
+CREATE INDEX IF NOT EXISTS idx_emb_corpus_scope_time
+    ON embeddings(corpus_id, scope_id, created_at_ts DESC);
+
+CREATE INDEX IF NOT EXISTS idx_emb_corpus_user_time
+    ON embeddings(corpus_id, user_id, created_at_ts DESC);
+
+CREATE INDEX IF NOT EXISTS idx_emb_corpus_org_time
+    ON embeddings(corpus_id, org_id, created_at_ts DESC);
+
+CREATE INDEX IF NOT EXISTS idx_emb_corpus_kind_time
+    ON embeddings(corpus_id, kind, created_at_ts DESC);
 """
 
 
 def _ensure_db(path: str) -> None:
     Path(path).parent.mkdir(parents=True, exist_ok=True)
-    conn = sqlite3.connect(path)
+    conn = sqlite3.connect(path, check_same_thread=False)
     try:
-        for stmt in SCHEMA.strip().split(";\n"):
+        cur = conn.cursor()
+        for stmt in SCHEMA.strip().split(";\n\n"):
             s = stmt.strip()
             if s:
-                conn.execute(s)
+                cur.execute(s)
         conn.commit()
     finally:
         conn.close()
 
 
+def _l2_normalize_rows(x: np.ndarray, eps: float = 1e-12) -> np.ndarray:
+    # x: (n, d)
+    norms = np.linalg.norm(x, axis=1, keepdims=True)
+    norms = np.maximum(norms, eps)
+    return x / norms
+
+
+def _l2_normalize_vec(x: np.ndarray, eps: float = 1e-12) -> np.ndarray:
+    n = float(np.linalg.norm(x))
+    if n < eps:
+        return x
+    return x / n
+
+
 class SQLiteVectorIndex(VectorIndex):
     """
     Simple SQLite-backed vector index.
-    Uses brute-force cosine similarity per corpus.
+
+    Baseline path uses brute-force cosine similarity over SQL-limited candidates. :contentReference[oaicite:1]{index=1}
+
+    Optional FAISS acceleration:
+      - If faiss is installed and enabled, maintains a per-corpus FAISS HNSW index on disk.
+      - Index is marked dirty on add/delete and rebuilt lazily on next search. NOTE: this can be slow for large corpora.
+      - This is a local index for small to medium workloads; for distributed or large-scale use cases, consider other backends.
+
+    Promoted fields you *may* pass in meta: :contentReference[oaicite:2]{index=2}
+      - scope_id, user_id, org_id, client_id, session_id
+      - run_id, graph_id, node_id
+      - kind, source
+      - created_at_ts (float UNIX timestamp)
     """
 
-    def __init__(self, root: str):
+    def __init__(
+        self,
+        root: str,
+        *,
+        use_faiss_if_available: bool = True,
+        faiss_m: int = 32,  # HNSW M
+        faiss_ef_search: int = 64,  # query-time accuracy/speed knob
+        faiss_ef_construction: int = 200,  # build-time accuracy/speed knob
+        faiss_probe_factor: int = 20,  # fetch k * factor candidates then post-filter
+        faiss_probe_min: int = 200,
+        faiss_probe_max: int = 5000,
+        brute_force_candidate_limit: int = 5000,
+    ):
         self.root = Path(root)
         self.root.mkdir(parents=True, exist_ok=True)
+
         self.db_path = str(self.root / "index.sqlite")
         _ensure_db(self.db_path)
 
+        # --- FAISS config ---
+        self._faiss_enabled = bool(use_faiss_if_available and faiss is not None)
+        self._faiss_dir = self.root / "faiss"
+        self._faiss_dir.mkdir(parents=True, exist_ok=True)
+
+        self._faiss_m = int(faiss_m)
+        self._faiss_ef_search = int(faiss_ef_search)
+        self._faiss_ef_construction = int(faiss_ef_construction)
+        self._faiss_probe_factor = int(faiss_probe_factor)
+        self._faiss_probe_min = int(faiss_probe_min)
+        self._faiss_probe_max = int(faiss_probe_max)
+
+        self._brute_force_candidate_limit = int(brute_force_candidate_limit)
+
+        self._faiss_lock = threading.RLock()
+        self._faiss_cache: dict[str, tuple[Any, list[str], int]] = {}
+        # cache: corpus_id -> (faiss_index, id_to_chunk_id, dim)
+        self._faiss_dirty: set[str] = set()
+
     def _connect(self) -> sqlite3.Connection:
-        # Each call gets its own connection: thread-safe with to_thread.
-        conn = sqlite3.connect(self.db_path, check_same_thread=False)
-        return conn
+        return sqlite3.connect(self.db_path, check_same_thread=False)
+
+    def _faiss_paths(self, corpus_id: str) -> tuple[Path, Path]:
+        safe = corpus_id.replace("/", "_")
+        return (self._faiss_dir / f"{safe}.index", self._faiss_dir / f"{safe}.meta.pkl")
+
+    def _mark_dirty(self, corpus_id: str) -> None:
+        if not self._faiss_enabled:
+            return
+        with self._faiss_lock:
+            self._faiss_dirty.add(corpus_id)
+            self._faiss_cache.pop(corpus_id, None)
+
+    def _build_faiss_index_from_db(self, corpus_id: str) -> tuple[Any, list[str], int]:
+        """
+        Build an HNSW cosine index for all vectors in a corpus.
+
+        We normalize vectors and use inner product (IP) => cosine similarity.
+        """
+        if not self._faiss_enabled:
+            raise RuntimeError("FAISS is not enabled/available.")
+
+        conn = self._connect()
+        try:
+            cur = conn.cursor()
+            cur.execute("SELECT chunk_id, vec FROM embeddings WHERE corpus_id=?", (corpus_id,))
+            rows = cur.fetchall()
+        finally:
+            conn.close()
+
+        if not rows:
+            # empty corpus index
+            dim = 0
+            index = None
+            return index, [], dim
+
+        # Infer dim from first vector
+        first_vec = np.frombuffer(rows[0][1], dtype=np.float32)
+        dim = int(first_vec.shape[0])
+
+        # Base HNSW (IP metric), wrapped with ID map so ids are stable ints
+        base = faiss.IndexHNSWFlat(dim, self._faiss_m, faiss.METRIC_INNER_PRODUCT)
+        base.hnsw.efConstruction = self._faiss_ef_construction
+        base.hnsw.efSearch = self._faiss_ef_search
+        index = faiss.IndexIDMap2(base)
+
+        id_to_chunk: list[str] = []
+        next_id = 0
+
+        # Add in batches to keep memory reasonable
+        batch_size = 2048
+        for i in range(0, len(rows), batch_size):
+            batch = rows[i : i + batch_size]
+            chunk_ids = [r[0] for r in batch]
+            mats = [np.frombuffer(r[1], dtype=np.float32) for r in batch]
+            x = np.stack(mats, axis=0).astype(np.float32, copy=False)
+            x = _l2_normalize_rows(x)
+
+            ids = np.arange(next_id, next_id + len(chunk_ids), dtype=np.int64)
+            index.add_with_ids(x, ids)
+
+            id_to_chunk.extend(chunk_ids)
+            next_id += len(chunk_ids)
+
+        return index, id_to_chunk, dim
+
+    def _ensure_faiss_ready(self, corpus_id: str) -> tuple[Any, list[str], int]:
+        if not self._faiss_enabled:
+            raise RuntimeError("FAISS is not enabled/available.")
+
+        with self._faiss_lock:
+            cached = self._faiss_cache.get(corpus_id)
+            if cached is not None and corpus_id not in self._faiss_dirty:
+                return cached
+
+            index_path, meta_path = self._faiss_paths(corpus_id)
+
+            # If not dirty and files exist, load from disk
+            if corpus_id not in self._faiss_dirty and index_path.exists() and meta_path.exists():
+                index = faiss.read_index(str(index_path))
+                with meta_path.open("rb") as f:
+                    meta = pickle.load(f)
+                id_to_chunk = meta["id_to_chunk"]
+                dim = int(meta["dim"])
+                # Ensure query-time params applied
+                try:
+                    index.index.hnsw.efSearch = self._faiss_ef_search  # type: ignore[attr-defined]
+                except Exception:
+                    import logging
+
+                    logger = logging.getLogger(__name__)
+                    logger.warning("Failed to set faiss efSearch parameter.")
+
+                self._faiss_cache[corpus_id] = (index, id_to_chunk, dim)
+                return index, id_to_chunk, dim
+
+            # Otherwise rebuild from DB
+            index, id_to_chunk, dim = self._build_faiss_index_from_db(corpus_id)
+
+            # Persist (if non-empty)
+            if index is not None:
+                faiss.write_index(index, str(index_path))
+                with meta_path.open("wb") as f:
+                    pickle.dump({"id_to_chunk": id_to_chunk, "dim": dim}, f)
+
+            self._faiss_dirty.discard(corpus_id)
+            self._faiss_cache[corpus_id] = (index, id_to_chunk, dim)
+            return index, id_to_chunk, dim
 
     async def add(
         self,
@@ -74,18 +271,72 @@ class SQLiteVectorIndex(VectorIndex):
                 for cid, vec, meta in zip(chunk_ids, vectors, metas, strict=True):
                     v = np.asarray(vec, dtype=np.float32)
                     norm = float(np.linalg.norm(v) + 1e-9)
+
+                    meta_json = json.dumps(meta, ensure_ascii=False)
+
+                    # promoted, optional
+                    scope_id = meta.get("scope_id")
+                    user_id = meta.get("user_id")
+                    org_id = meta.get("org_id")
+                    client_id = meta.get("client_id")
+                    session_id = meta.get("session_id")
+                    run_id = meta.get("run_id")
+                    graph_id = meta.get("graph_id")
+                    node_id = meta.get("node_id")
+                    kind = meta.get("kind")
+                    source = meta.get("source")
+                    created_at_ts = meta.get("created_at_ts")
+
                     cur.execute(
                         "REPLACE INTO chunks(corpus_id,chunk_id,meta_json) VALUES(?,?,?)",
-                        (corpus_id, cid, json.dumps(meta, ensure_ascii=False)),
+                        (corpus_id, cid, meta_json),
                     )
                     cur.execute(
-                        "REPLACE INTO embeddings(corpus_id,chunk_id,vec,norm) VALUES(?,?,?,?)",
-                        (corpus_id, cid, v.tobytes(), norm),
+                        """
+                        REPLACE INTO embeddings(
+                            corpus_id,
+                            chunk_id,
+                            vec,
+                            norm,
+                            scope_id,
+                            user_id,
+                            org_id,
+                            client_id,
+                            session_id,
+                            run_id,
+                            graph_id,
+                            node_id,
+                            kind,
+                            source,
+                            created_at_ts
+                        )
+                        VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)
+                        """,
+                        (
+                            corpus_id,
+                            cid,
+                            v.tobytes(),
+                            norm,
+                            scope_id,
+                            user_id,
+                            org_id,
+                            client_id,
+                            session_id,
+                            run_id,
+                            graph_id,
+                            node_id,
+                            kind,
+                            source,
+                            created_at_ts,
+                        ),
                     )
                 conn.commit()
             finally:
                 conn.close()
 
+            # Mark FAISS corpus index dirty (lazy rebuild on next search)
+            self._mark_dirty(corpus_id)
+
         await asyncio.to_thread(_add_sync)
 
     async def delete(self, corpus_id: str, chunk_ids: list[str] | None = None) -> None:
@@ -110,6 +361,9 @@ class SQLiteVectorIndex(VectorIndex):
             finally:
                 conn.close()
 
+            # Mark FAISS corpus index dirty (lazy rebuild on next search)
+            self._mark_dirty(corpus_id)
+
         await asyncio.to_thread(_delete_sync)
 
     async def list_chunks(self, corpus_id: str) -> list[str]:
@@ -136,52 +390,267 @@ class SQLiteVectorIndex(VectorIndex):
 
         return await asyncio.to_thread(_list_sync)
 
+    def _passes_where(self, where: dict[str, Any], row: dict[str, Any]) -> bool:
+        # where only applies to promoted fields in your SQL path :contentReference[oaicite:3]{index=3}
+        # Here, we enforce the same semantics post-hoc.
+        for k, v in where.items():
+            if v is None:
+                continue
+            if row.get(k) != v:
+                return False
+        return True
+
+    def _fetch_rows_for_chunk_ids(
+        self,
+        conn: sqlite3.Connection,
+        corpus_id: str,
+        chunk_ids: list[str],
+    ) -> dict[str, tuple[str, float | None]]:
+        """
+        Returns {chunk_id: (meta_json, created_at_ts)} for given chunk_ids.
+        Batched to avoid SQLite parameter limits.
+        """
+        out: dict[str, tuple[str, float | None]] = {}
+        cur = conn.cursor()
+
+        # SQLite default param limit is often 999, keep headroom
+        batch_size = 900
+        for i in range(0, len(chunk_ids), batch_size):
+            b = chunk_ids[i : i + batch_size]
+            placeholders = ",".join("?" for _ in b)
+            sql = f"""
+                SELECT e.chunk_id, c.meta_json, e.created_at_ts
+                FROM embeddings e
+                JOIN chunks c
+                ON e.corpus_id = c.corpus_id AND e.chunk_id = c.chunk_id
+                WHERE e.corpus_id = ?
+                AND e.chunk_id IN ({placeholders})
+            """
+            cur.execute(sql, [corpus_id, *b])
+            for cid, meta_json, created_at_ts in cur.fetchall():
+                out[str(cid)] = (str(meta_json), created_at_ts)
+        return out
+
+    def _search_bruteforce_sync(
+        self,
+        corpus_id: str,
+        q: np.ndarray,
+        k: int,
+        where: dict[str, Any],
+        max_candidates: int | None,
+        created_at_min: float | None,
+        created_at_max: float | None,
+    ) -> list[dict[str, Any]]:
+        # This is your original SQL-candidate path (kept as a fallback). :contentReference[oaicite:4]{index=4}
+        qn = float(np.linalg.norm(q) + 1e-9)
+
+        conn = self._connect()
+        try:
+            cur = conn.cursor()
+
+            sql = """
+                SELECT e.chunk_id, e.vec, e.norm, c.meta_json
+                FROM embeddings e
+                JOIN chunks c
+                ON e.corpus_id = c.corpus_id AND e.chunk_id = c.chunk_id
+                WHERE e.corpus_id=?
+            """
+            params: list[Any] = [corpus_id]
+
+            promoted_cols = {
+                "scope_id",
+                "user_id",
+                "org_id",
+                "client_id",
+                "session_id",
+                "run_id",
+                "graph_id",
+                "node_id",
+                "kind",
+                "source",
+            }
+
+            for key, val in where.items():
+                if val is None:
+                    continue
+                if key in promoted_cols:
+                    sql += f" AND e.{key} = ?"
+                    params.append(val)
+
+            if created_at_min is not None:
+                sql += " AND e.created_at_ts >= ?"
+                params.append(created_at_min)
+            if created_at_max is not None:
+                sql += " AND e.created_at_ts <= ?"
+                params.append(created_at_max)
+
+            candidate_limit = max_candidates or self._brute_force_candidate_limit
+            sql += " ORDER BY e.created_at_ts DESC"
+            sql += " LIMIT ?"
+            params.append(candidate_limit)
+
+            cur.execute(sql, params)
+            rows = cur.fetchall()
+        finally:
+            conn.close()
+
+        # Minor speedup: avoid json.loads until after top-k
+        scored: list[tuple[float, str, str]] = []
+        for chunk_id, vec_bytes, norm, meta_json in rows:
+            v = np.frombuffer(vec_bytes, dtype=np.float32)
+            score = float(np.dot(q, v) / (qn * (norm or 1e-9)))
+            scored.append((score, str(chunk_id), str(meta_json)))
+
+        scored.sort(key=lambda x: x[0], reverse=True)
+        top = scored[:k]
+
+        out: list[dict[str, Any]] = []
+        for score, chunk_id, meta_json in top:
+            out.append({"chunk_id": chunk_id, "score": score, "meta": json.loads(meta_json)})
+        return out
+
+    def _search_faiss_sync(
+        self,
+        corpus_id: str,
+        q: np.ndarray,
+        k: int,
+        where: dict[str, Any],
+        created_at_min: float | None,
+        created_at_max: float | None,
+        max_candidates: int | None,
+    ) -> list[dict[str, Any]]:
+        # If max_candidates is set very small, caller probably expects strict recency-bounded behavior.
+        # In that case, keep the old semantics.
+        if max_candidates is not None and max_candidates <= self._brute_force_candidate_limit:
+            return self._search_bruteforce_sync(
+                corpus_id=corpus_id,
+                q=q,
+                k=k,
+                where=where,
+                max_candidates=max_candidates,
+                created_at_min=created_at_min,
+                created_at_max=created_at_max,
+            )
+
+        index, id_to_chunk, dim = self._ensure_faiss_ready(corpus_id)
+        if index is None or dim <= 0 or not id_to_chunk:
+            return []
+
+        if q.shape[0] != dim:
+            # Dim mismatch: fall back to brute-force rather than throwing.
+            return self._search_bruteforce_sync(
+                corpus_id=corpus_id,
+                q=q,
+                k=k,
+                where=where,
+                max_candidates=max_candidates,
+                created_at_min=created_at_min,
+                created_at_max=created_at_max,
+            )
+
+        qn = _l2_normalize_vec(q.astype(np.float32, copy=False))
+        qn = qn.reshape(1, -1)
+
+        # Probe progressively deeper until we have k results that pass filters
+        probe = max(self._faiss_probe_min, k * self._faiss_probe_factor)
+        probe = min(probe, self._faiss_probe_max)
+
+        conn = self._connect()
+        try:
+            while True:
+                scores, ids = index.search(qn, probe)
+                ids0 = ids[0]
+                scores0 = scores[0]
+
+                # Map to chunk_ids in rank order
+                ranked_chunk_ids: list[str] = []
+                ranked_scores: list[float] = []
+                for fid, sc in zip(ids0, scores0, strict=False):
+                    if fid < 0:
+                        continue
+                    if fid >= len(id_to_chunk):
+                        continue
+                    ranked_chunk_ids.append(id_to_chunk[int(fid)])
+                    ranked_scores.append(float(sc))
+
+                if not ranked_chunk_ids:
+                    return []
+
+                # Fetch metas/timestamps for these candidates in batch
+                row_map = self._fetch_rows_for_chunk_ids(conn, corpus_id, ranked_chunk_ids)
+
+                out: list[dict[str, Any]] = []
+                for cid, sc in zip(ranked_chunk_ids, ranked_scores, strict=True):
+                    tup = row_map.get(cid)
+                    if tup is None:
+                        continue
+                    meta_json, created_at_ts = tup
+                    meta = json.loads(meta_json)
+
+                    # Apply where/time filters post-hoc
+                    if where and not self._passes_where(where, meta):
+                        continue
+                    if created_at_min is not None:  # noqa: SIM102
+                        if created_at_ts is None or float(created_at_ts) < float(created_at_min):
+                            continue
+                    if created_at_max is not None:  # noqa: SIM102
+                        if created_at_ts is None or float(created_at_ts) > float(created_at_max):
+                            continue
+
+                    out.append({"chunk_id": cid, "score": sc, "meta": meta})
+                    if len(out) >= k:
+                        return out
+
+                # Not enough after filtering -> probe deeper or fall back
+                if probe >= self._faiss_probe_max:
+                    # If filters are too tight, FAISS post-filtering may not find enough.
+                    # Fall back to SQL-candidate brute-force which is exact under filters.
+                    return self._search_bruteforce_sync(
+                        corpus_id=corpus_id,
+                        q=q,
+                        k=k,
+                        where=where,
+                        max_candidates=max_candidates,
+                        created_at_min=created_at_min,
+                        created_at_max=created_at_max,
+                    )
+
+                probe = min(self._faiss_probe_max, probe * 2)
+        finally:
+            conn.close()
+
     async def search(
         self,
         corpus_id: str,
         query_vec: list[float],
         k: int,
+        where: dict[str, Any] | None = None,
+        max_candidates: int | None = None,
+        created_at_min: float | None = None,
+        created_at_max: float | None = None,
     ) -> list[dict[str, Any]]:
         q = np.asarray(query_vec, dtype=np.float32)
-        qn = float(np.linalg.norm(q) + 1e-9)
+        where = where or {}
 
         def _search_sync() -> list[dict[str, Any]]:
-            conn = self._connect()
-            try:
-                cur = conn.cursor()
-                cur.execute(
-                    """
-                    SELECT e.chunk_id, e.vec, e.norm, c.meta_json
-                    FROM embeddings e
-                    JOIN chunks c
-                      ON e.corpus_id = c.corpus_id AND e.chunk_id = c.chunk_id
-                    WHERE e.corpus_id=?
-                    """,
-                    (corpus_id,),
-                )
-                rows = cur.fetchall()
-            finally:
-                conn.close()
-
-            scored: list[tuple[float, str, dict[str, Any]]] = []
-            for chunk_id, vec_bytes, norm, meta_json in rows:
-                v = np.frombuffer(vec_bytes, dtype=np.float32)
-                score = float(np.dot(q, v) / (qn * norm))
-                meta = json.loads(meta_json)
-                scored.append((score, chunk_id, meta))
-
-            scored.sort(key=lambda x: x[0], reverse=True)
-            top = scored[:k]
-
-            out: list[dict[str, Any]] = []
-            for score, chunk_id, meta in top:
-                out.append(
-                    {
-                        "chunk_id": chunk_id,
-                        "score": score,
-                        "meta": meta,
-                    }
+            if self._faiss_enabled:
+                return self._search_faiss_sync(
+                    corpus_id=corpus_id,
+                    q=q,
+                    k=k,
+                    where=where,
+                    created_at_min=created_at_min,
+                    created_at_max=created_at_max,
+                    max_candidates=max_candidates,
                 )
-            return out
+            return self._search_bruteforce_sync(
+                corpus_id=corpus_id,
+                q=q,
+                k=k,
+                where=where,
+                max_candidates=max_candidates,
+                created_at_min=created_at_min,
+                created_at_max=created_at_max,
+            )
 
         return await asyncio.to_thread(_search_sync)