@remnic/core 1.1.12 → 1.1.14

package/scripts/faiss_index.py

@@ -0,0 +1,816 @@
+#!/usr/bin/env python3
+"""FAISS conversation index sidecar.
+
+JSON-in/JSON-out CLI used by src/conversation-index/faiss-adapter.ts.
+"""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import os
+import subprocess
+import sys
+import time
+import uuid
+from pathlib import Path
+from typing import Any
+
+MODEL_CACHE: dict[str, Any] = {}
+LOCK_OWNERS: dict[str, str] = {}
+HASH_EMBED_DIM = 128
+LOCK_TIMEOUT_SECONDS = 10.0
+LOCK_STALE_SECONDS = 120.0
+MANIFEST_VERSION = 1
+MODEL_ID_ALIASES = {
+    "text-embedding-3-small": "sentence-transformers/all-MiniLM-L6-v2",
+    "text-embedding-3-large": "sentence-transformers/all-mpnet-base-v2",
+    "text-embedding-ada-002": "sentence-transformers/all-MiniLM-L6-v2",
+}
+
+
+class SidecarError(Exception):
+    pass
+
+
+class DependencyError(SidecarError):
+    pass
+
+
+def emit(payload: dict[str, Any]) -> None:
+    sys.stdout.write(json.dumps(payload, separators=(",", ":"), ensure_ascii=False))
+    sys.stdout.flush()
+
+
+def read_payload() -> dict[str, Any]:
+    raw = sys.stdin.read()
+    if not raw.strip():
+        raise SidecarError("empty stdin payload")
+    try:
+        payload = json.loads(raw)
+    except json.JSONDecodeError as exc:
+        raise SidecarError(f"invalid JSON payload: {exc}") from exc
+    if not isinstance(payload, dict):
+        raise SidecarError("payload must be a JSON object")
+    return payload
+
+
+def ensure_index_dir(index_path: str) -> Path:
+    if not isinstance(index_path, str) or not index_path.strip():
+        raise SidecarError("indexPath is required")
+    path = Path(index_path)
+    path.mkdir(parents=True, exist_ok=True)
+    return path
+
+
+def metadata_file(index_dir: Path) -> Path:
+    return index_dir / "metadata.jsonl"
+
+
+def index_file(index_dir: Path) -> Path:
+    return index_dir / "index.faiss"
+
+
+def manifest_file(index_dir: Path) -> Path:
+    return index_dir / "manifest.json"
+
+
+def read_metadata(path: Path) -> list[dict[str, Any]]:
+    if not path.exists():
+        return []
+
+    rows: list[dict[str, Any]] = []
+    for line in path.read_text(encoding="utf-8").splitlines():
+        line = line.strip()
+        if not line:
+            continue
+        try:
+            row = json.loads(line)
+        except json.JSONDecodeError:
+            continue
+        if not isinstance(row, dict):
+            continue
+        row_id = row.get("id")
+        text = row.get("text")
+        if not isinstance(row_id, str) or not row_id:
+            continue
+        if not isinstance(text, str):
+            continue
+        rows.append(
+            {
+                "id": row_id,
+                "sessionKey": row.get("sessionKey") if isinstance(row.get("sessionKey"), str) else "",
+                "text": text,
+                "startTs": row.get("startTs") if isinstance(row.get("startTs"), str) else "",
+                "endTs": row.get("endTs") if isinstance(row.get("endTs"), str) else "",
+            }
+        )
+    return rows
+
+
+def write_metadata(path: Path, rows: list[dict[str, Any]]) -> None:
+    tmp = path.with_suffix(".jsonl.tmp")
+    with tmp.open("w", encoding="utf-8") as handle:
+        for row in rows:
+            handle.write(json.dumps(row, separators=(",", ":"), ensure_ascii=False))
+            handle.write("\n")
+    os.replace(tmp, path)
+
+
+def read_manifest(path: Path) -> dict[str, Any] | None:
+    if not path.exists():
+        return None
+    try:
+        raw = json.loads(path.read_text(encoding="utf-8"))
+    except Exception:
+        return None
+    return raw if isinstance(raw, dict) else None
+
+
+def write_manifest(path: Path, manifest: dict[str, Any]) -> None:
+    tmp = path.with_suffix(".json.tmp")
+    tmp.write_text(
+        json.dumps(manifest, separators=(",", ":"), ensure_ascii=False),
+        encoding="utf-8",
+    )
+    os.replace(tmp, path)
+
+
+def load_vector_dependencies() -> tuple[Any, Any]:
+    try:
+        import numpy as np  # type: ignore
+        import faiss  # type: ignore
+    except Exception as exc:
+        raise DependencyError(f"missing faiss dependencies: {exc}") from exc
+    return np, faiss
+
+
+def sentence_transformers_enabled() -> bool:
+    if "REMNIC_FAISS_ENABLE_ST" in os.environ:
+        value = os.environ["REMNIC_FAISS_ENABLE_ST"]
+    else:
+        value = os.environ.get("ENGRAM_FAISS_ENABLE_ST", "")
+    value = value.strip().lower()
+    return value in ("1", "true", "yes", "on")
+
+
+def normalize_model_id(model_id: str) -> str:
+    cleaned = (model_id or "").strip()
+    if not cleaned:
+        cleaned = "sentence-transformers/all-MiniLM-L6-v2"
+    resolved = MODEL_ID_ALIASES.get(cleaned, cleaned)
+    if resolved in ("__hash__", "hash"):
+        return "__hash__"
+    if not sentence_transformers_enabled():
+        return "__hash__"
+    return resolved
+
+
+def get_embedder(model_id: str) -> Any:
+    resolved_model_id = normalize_model_id(model_id)
+    if resolved_model_id == "__hash__":
+        return None
+    if resolved_model_id in MODEL_CACHE:
+        return MODEL_CACHE[resolved_model_id]
+    try:
+        from sentence_transformers import SentenceTransformer  # type: ignore
+    except Exception as exc:
+        raise DependencyError(f"missing sentence-transformers dependency: {exc}") from exc
+    MODEL_CACHE[resolved_model_id] = SentenceTransformer(resolved_model_id)
+    return MODEL_CACHE[resolved_model_id]
+
+
+def embed_with_hash(texts: list[str], np: Any) -> Any:
+    vectors = np.zeros((len(texts), HASH_EMBED_DIM), dtype="float32")
+    for row_index, text in enumerate(texts):
+        digest = hashlib.sha256(text.encode("utf-8")).digest()
+        for byte_index in range(HASH_EMBED_DIM):
+            vectors[row_index, byte_index] = (digest[byte_index % len(digest)] / 255.0) - 0.5
+    return vectors
+
+
+def embed_texts(texts: list[str], model_id: str) -> tuple[Any, Any, Any]:
+    np, faiss = load_vector_dependencies()
+    embedder = get_embedder(model_id)
+    if embedder is None:
+        arr = embed_with_hash(texts, np)
+    else:
+        vectors = embedder.encode(
+            texts,
+            normalize_embeddings=False,
+            convert_to_numpy=True,
+            show_progress_bar=False,
+        )
+        arr = np.asarray(vectors, dtype="float32")
+    if arr.ndim == 1:
+        arr = arr.reshape(1, -1)
+    if arr.shape[0] > 0:
+        faiss.normalize_L2(arr)
+    return arr, np, faiss
+
+
+def write_index(path: Path, vectors: Any, faiss: Any) -> None:
+    dim = int(vectors.shape[1])
+    index = faiss.IndexFlatIP(dim)
+    if int(vectors.shape[0]) > 0:
+        index.add(vectors)
+    tmp = path.with_suffix(".faiss.tmp")
+    faiss.write_index(index, str(tmp))
+    os.replace(tmp, path)
+
+
+def resolve_vector_dimension(model_id: str) -> int:
+    probe, _np, _faiss = embed_texts([""], model_id)
+    return int(probe.shape[1])
+
+
+def build_empty_vectors(model_id: str) -> tuple[Any, Any]:
+    np, faiss = load_vector_dependencies()
+    dim = resolve_vector_dimension(model_id)
+    vectors = np.zeros((0, dim), dtype="float32")
+    return vectors, faiss
+
+
+def build_manifest(
+    model_id: str,
+    vector_dim: int,
+    chunk_count: int,
+    *,
+    generated_at: str | None = None,
+    last_successful_rebuild_at: str | None = None,
+) -> dict[str, Any]:
+    normalized_model_id = normalize_model_id(model_id)
+    now_iso = generated_at or time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
+    return {
+        "version": MANIFEST_VERSION,
+        "modelId": model_id,
+        "normalizedModelId": normalized_model_id,
+        "dimension": int(vector_dim),
+        "chunkCount": int(chunk_count),
+        "updatedAt": now_iso,
+        "lastSuccessfulRebuildAt": last_successful_rebuild_at or now_iso,
+    }
+
+
+def validate_index_manifest(
+    manifest: dict[str, Any] | None,
+    *,
+    requested_model_id: str,
+    actual_dimension: int | None = None,
+    expected_dimension: int | None = None,
+) -> dict[str, Any]:
+    if manifest is None:
+        raise SidecarError("missing index manifest; rebuild the FAISS conversation index")
+
+    version = manifest.get("version")
+    if not isinstance(version, int) or isinstance(version, bool) or version != MANIFEST_VERSION:
+        raise SidecarError("unsupported index manifest version; rebuild the FAISS conversation index")
+
+    normalized_manifest_model_id = manifest.get("normalizedModelId")
+    if not isinstance(normalized_manifest_model_id, str) or not normalized_manifest_model_id:
+        raise SidecarError("index manifest missing normalized model id; rebuild the FAISS conversation index")
+
+    requested_normalized_model_id = normalize_model_id(requested_model_id)
+    if normalized_manifest_model_id != requested_normalized_model_id:
+        raise SidecarError(
+            "index model mismatch "
+            f"(index={normalized_manifest_model_id}, query={requested_normalized_model_id}); "
+            "rebuild the FAISS conversation index"
+        )
+
+    manifest_dimension = manifest.get("dimension")
+    if not isinstance(manifest_dimension, int) or isinstance(manifest_dimension, bool) or manifest_dimension <= 0:
+        raise SidecarError("index manifest missing vector dimension; rebuild the FAISS conversation index")
+
+    chunk_count = manifest.get("chunkCount")
+    if not isinstance(chunk_count, int) or isinstance(chunk_count, bool) or chunk_count < 0:
+        raise SidecarError("index manifest missing chunk count; rebuild the FAISS conversation index")
+
+    if actual_dimension is not None and manifest_dimension != int(actual_dimension):
+        raise SidecarError(
+            f"index dimension mismatch (manifest={manifest_dimension}, index={int(actual_dimension)}); "
+            "rebuild the FAISS conversation index"
+        )
+
+    if expected_dimension is not None and manifest_dimension != int(expected_dimension):
+        raise SidecarError(
+            f"index dimension mismatch (manifest={manifest_dimension}, query={int(expected_dimension)}); "
+            "rebuild the FAISS conversation index"
+        )
+
+    return {
+        "version": version,
+        "modelId": manifest.get("modelId") if isinstance(manifest.get("modelId"), str) else "",
+        "normalizedModelId": normalized_manifest_model_id,
+        "dimension": manifest_dimension,
+        "chunkCount": chunk_count,
+        "updatedAt": manifest.get("updatedAt") if isinstance(manifest.get("updatedAt"), str) else "",
+        "lastSuccessfulRebuildAt": (
+            manifest.get("lastSuccessfulRebuildAt")
+            if isinstance(manifest.get("lastSuccessfulRebuildAt"), str)
+            else ""
+        ),
+    }
+
+
+def validate_artifact_counts(index: Any, rows: list[dict[str, Any]], manifest: dict[str, Any]) -> None:
+    index_count = int(getattr(index, "ntotal", -1))
+    metadata_count = len(rows)
+    manifest_count = manifest["chunkCount"]
+
+    if index_count != metadata_count or index_count != int(manifest_count):
+        raise SidecarError(
+            "index artifact count mismatch "
+            f"(index={index_count}, metadata={metadata_count}, manifest={int(manifest_count)}); "
+            "rebuild the FAISS conversation index"
+        )
+
+
+def parse_chunks(payload: dict[str, Any]) -> list[dict[str, Any]]:
+    raw_chunks = payload.get("chunks")
+    if not isinstance(raw_chunks, list):
+        raise SidecarError("chunks must be an array")
+    chunks: list[dict[str, Any]] = []
+    for item in raw_chunks:
+        if not isinstance(item, dict):
+            continue
+        chunk_id = item.get("id")
+        text = item.get("text")
+        if not isinstance(chunk_id, str) or not chunk_id:
+            continue
+        if not isinstance(text, str):
+            continue
+        chunks.append(
+            {
+                "id": chunk_id,
+                "sessionKey": item.get("sessionKey") if isinstance(item.get("sessionKey"), str) else "",
+                "text": text,
+                "startTs": item.get("startTs") if isinstance(item.get("startTs"), str) else "",
+                "endTs": item.get("endTs") if isinstance(item.get("endTs"), str) else "",
+            }
+        )
+    return chunks
+
+
+def metadata_row_key(row: dict[str, Any]) -> tuple[str, str]:
+    session_key = (
+        row.get("sessionKey") if isinstance(row.get("sessionKey"), str) else ""
+    )
+    row_id = row.get("id") if isinstance(row.get("id"), str) else ""
+    return (session_key, row_id)
+
+
+def merge_rows(existing: list[dict[str, Any]], updates: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    by_key: dict[tuple[str, str], dict[str, Any]] = {
+        metadata_row_key(row): row for row in existing
+    }
+    order = [metadata_row_key(row) for row in existing]
+    for update in updates:
+        update_key = metadata_row_key(update)
+        if update_key not in by_key:
+            order.append(update_key)
+        by_key[update_key] = update
+    return [by_key[row_key] for row_key in order]
+
+
+def lock_owner_key(lock_path: Path) -> str:
+    return str(lock_path.resolve())
+
+
+def make_lock_owner_token() -> str:
+    return f"{os.getpid()}:{uuid.uuid4().hex}"
+
+
+def read_lock_contents(lock_path: Path) -> str | None:
+    try:
+        raw = lock_path.read_text(encoding="utf-8").strip()
+    except Exception:
+        return None
+    if not raw:
+        return None
+    return raw
+
+
+def read_lock_owner_pid(lock_path: Path) -> int | None:
+    raw = read_lock_contents(lock_path)
+    if raw is None:
+        return None
+    pid_raw = raw.split(":", 1)[0].strip()
+    try:
+        pid = int(pid_raw)
+    except ValueError:
+        return None
+    return pid if pid > 0 else None
+
+
+def lock_stat_matches(current_stat: os.stat_result, observed_stat: os.stat_result) -> bool:
+    if current_stat.st_size != observed_stat.st_size:
+        return False
+    if current_stat.st_mtime_ns != observed_stat.st_mtime_ns:
+        return False
+    if getattr(current_stat, "st_ino", 0) and getattr(observed_stat, "st_ino", 0):
+        return current_stat.st_ino == observed_stat.st_ino
+    return True
+
+
+def unlink_lock_if_unchanged(
+    lock_path: Path,
+    observed_owner_token: str | None,
+    observed_stat: os.stat_result,
+) -> bool:
+    try:
+        current_stat = lock_path.stat()
+    except FileNotFoundError:
+        return False
+    current_owner_token = read_lock_contents(lock_path)
+    if current_owner_token != observed_owner_token:
+        return False
+    if not lock_stat_matches(current_stat, observed_stat):
+        return False
+    try:
+        lock_path.unlink()
+    except FileNotFoundError:
+        return False
+    return True
+
+
+def is_process_alive(pid: int) -> bool:
+    if pid <= 0:
+        return False
+
+    if os.name == "nt":
+        try:
+            probe = subprocess.run(
+                ["tasklist", "/FI", f"PID eq {pid}", "/FO", "CSV", "/NH"],
+                capture_output=True,
+                text=True,
+                timeout=2,
+            )
+        except Exception:
+            return False
+        output = probe.stdout.strip()
+        if not output:
+            return False
+        if output.startswith("INFO:"):
+            return False
+        return f'"{pid}"' in output
+
+    try:
+        os.kill(pid, 0)
+        return True
+    except ProcessLookupError:
+        return False
+    except PermissionError:
+        return True
+    except OSError:
+        return False
+
+
+def acquire_lock(index_dir: Path, lock_name: str) -> Path:
+    lock_path = index_dir / lock_name
+    lock_owner_token = make_lock_owner_token()
+    lock_label = lock_name.strip(".")
+    if lock_label.endswith(".lock"):
+        lock_label = lock_label[: -len(".lock")]
+    lock_label = lock_label.replace(".", " ")
+    deadline = time.monotonic() + LOCK_TIMEOUT_SECONDS
+
+    while True:
+        try:
+            fd = os.open(str(lock_path), os.O_CREAT | os.O_EXCL | os.O_WRONLY)
+            with os.fdopen(fd, "w", encoding="utf-8") as handle:
+                handle.write(lock_owner_token)
+            LOCK_OWNERS[lock_owner_key(lock_path)] = lock_owner_token
+            return lock_path
+        except FileExistsError:
+            try:
+                observed_stat = lock_path.stat()
+            except FileNotFoundError:
+                continue
+
+            age = time.time() - observed_stat.st_mtime
+            observed_owner_token = read_lock_contents(lock_path)
+            owner_pid = read_lock_owner_pid(lock_path)
+            owner_alive = is_process_alive(owner_pid) if owner_pid is not None else False
+
+            if age > LOCK_STALE_SECONDS and not owner_alive:
+                unlink_lock_if_unchanged(lock_path, observed_owner_token, observed_stat)
+                continue
+
+            if time.monotonic() >= deadline:
+                raise SidecarError(f"timed out waiting for FAISS {lock_label} lock")
+            time.sleep(0.05)
+
+
+def acquire_index_lock(index_dir: Path) -> Path:
+    return acquire_lock(index_dir, ".index.lock")
+
+
+def acquire_writer_lock(index_dir: Path) -> Path:
+    return acquire_lock(index_dir, ".writer.lock")
+
+
+def release_lock(lock_path: Path) -> None:
+    lock_owner_token = LOCK_OWNERS.pop(lock_owner_key(lock_path), None)
+    if lock_owner_token is None:
+        return
+    try:
+        observed_stat = lock_path.stat()
+    except FileNotFoundError:
+        return
+    unlink_lock_if_unchanged(lock_path, lock_owner_token, observed_stat)
+
+
+def release_index_lock(lock_path: Path) -> None:
+    release_lock(lock_path)
+
+
+def run_upsert(payload: dict[str, Any]) -> dict[str, Any]:
+    model_id = payload.get("modelId")
+    if not isinstance(model_id, str) or not model_id:
+        raise SidecarError("modelId is required")
+
+    index_dir = ensure_index_dir(str(payload.get("indexPath", "")))
+    chunks = parse_chunks(payload)
+
+    if not chunks:
+        return {"ok": True, "upserted": 0}
+
+    meta_path = metadata_file(index_dir)
+    idx_path = index_file(index_dir)
+    manifest_path = manifest_file(index_dir)
+
+    writer_lock_path = acquire_writer_lock(index_dir)
+    try:
+        existing = read_metadata(meta_path)
+        existing_manifest = read_manifest(manifest_path)
+        merged = merge_rows(existing, chunks)
+
+        texts = [row["text"] for row in merged]
+        vectors, _np, faiss = embed_texts(texts, model_id)
+
+        lock_path = acquire_index_lock(index_dir)
+        try:
+            # Commit FAISS index first; metadata follows so we never point at missing vectors.
+            write_index(idx_path, vectors, faiss)
+            write_metadata(meta_path, merged)
+            preserved_rebuild_at = (
+                existing_manifest.get("lastSuccessfulRebuildAt")
+                if isinstance(existing_manifest, dict)
+                and isinstance(existing_manifest.get("lastSuccessfulRebuildAt"), str)
+                and existing_manifest.get("lastSuccessfulRebuildAt")
+                else None
+            )
+            write_manifest(
+                manifest_path,
+                build_manifest(
+                    model_id,
+                    int(vectors.shape[1]),
+                    len(merged),
+                    last_successful_rebuild_at=preserved_rebuild_at,
+                ),
+            )
+        finally:
+            release_index_lock(lock_path)
+    finally:
+        release_lock(writer_lock_path)
+
+    return {"ok": True, "upserted": len(chunks)}
+
+
+def run_rebuild(payload: dict[str, Any]) -> dict[str, Any]:
+    model_id = payload.get("modelId")
+    if not isinstance(model_id, str) or not model_id:
+        raise SidecarError("modelId is required")
+
+    index_dir = ensure_index_dir(str(payload.get("indexPath", "")))
+    chunks = parse_chunks(payload)
+
+    writer_lock_path = acquire_writer_lock(index_dir)
+    try:
+        if chunks:
+            texts = [row["text"] for row in chunks]
+            vectors, _np, faiss = embed_texts(texts, model_id)
+            chunk_count = len(chunks)
+        else:
+            vectors, faiss = build_empty_vectors(model_id)
+            chunk_count = 0
+
+        lock_path = acquire_index_lock(index_dir)
+        try:
+            meta_path = metadata_file(index_dir)
+            idx_path = index_file(index_dir)
+            manifest_path = manifest_file(index_dir)
+
+            write_index(idx_path, vectors, faiss)
+            write_metadata(meta_path, chunks)
+            write_manifest(manifest_path, build_manifest(model_id, int(vectors.shape[1]), chunk_count))
+        finally:
+            release_index_lock(lock_path)
+    finally:
+        release_lock(writer_lock_path)
+
+    return {"ok": True, "rebuilt": len(chunks)}
+
+
+def run_search(payload: dict[str, Any]) -> dict[str, Any]:
+    model_id = payload.get("modelId")
+    query = payload.get("query")
+    top_k = payload.get("topK")
+    if not isinstance(model_id, str) or not model_id:
+        raise SidecarError("modelId is required")
+    if not isinstance(query, str) or not query.strip():
+        raise SidecarError("query is required")
+    if not isinstance(top_k, int) or isinstance(top_k, bool) or top_k <= 0:
+        raise SidecarError("topK must be a positive integer")
+
+    index_dir = ensure_index_dir(str(payload.get("indexPath", "")))
+    meta_path = metadata_file(index_dir)
+    idx_path = index_file(index_dir)
+    manifest_path = manifest_file(index_dir)
+
+    lock_path = acquire_index_lock(index_dir)
+    try:
+        has_index = idx_path.exists()
+        has_metadata = meta_path.exists()
+        has_manifest = manifest_path.exists()
+        if not has_index and not has_metadata and not has_manifest:
+            return {"ok": True, "results": []}
+        if not has_index or not has_metadata or not has_manifest:
+            raise SidecarError("conversation index artifacts incomplete; rebuild the FAISS conversation index")
+
+        rows = read_metadata(meta_path)
+        _np, faiss = load_vector_dependencies()
+        index = faiss.read_index(str(idx_path))
+        manifest = validate_index_manifest(
+            read_manifest(manifest_path),
+            requested_model_id=model_id,
+            actual_dimension=int(index.d),
+        )
+        validate_artifact_counts(index, rows, manifest)
+    finally:
+        release_index_lock(lock_path)
+
+    if not rows:
+        return {"ok": True, "results": []}
+
+    query_vector, _np2, _faiss2 = embed_texts([query], model_id)
+    query_dimension = int(query_vector.shape[1])
+    if int(manifest["dimension"]) != query_dimension:
+        raise SidecarError(
+            f"index dimension mismatch (manifest={int(manifest['dimension'])}, query={query_dimension}); "
+            "rebuild the FAISS conversation index"
+        )
+
+    distances, indices = index.search(query_vector, top_k)
+    results: list[dict[str, Any]] = []
+    for score, idx in zip(distances[0], indices[0]):
+        idx_i = int(idx)
+        if idx_i < 0 or idx_i >= len(rows):
+            continue
+        row = rows[idx_i]
+        results.append(
+            {
+                "path": row["id"],
+                "snippet": row["text"][:280],
+                "score": float(score),
+            }
+        )
+
+    return {"ok": True, "results": results}
+
+
+def build_health_response(payload: dict[str, Any], *, include_metadata: bool = False) -> dict[str, Any]:
+    index_dir = ensure_index_dir(str(payload.get("indexPath", "")))
+    meta_path = metadata_file(index_dir)
+    idx_path = index_file(index_dir)
+    manifest_path = manifest_file(index_dir)
+
+    status = "ok"
+    error = ""
+    model_id = normalize_model_id(str(payload.get("modelId", "")))
+    manifest_details: dict[str, Any] | None = None
+    metadata_details: dict[str, Any] | None = None
+
+    try:
+        load_vector_dependencies()
+        if model_id != "__hash__":
+            try:
+                import sentence_transformers  # type: ignore # noqa: F401
+            except Exception as exc:
+                raise DependencyError(f"missing sentence-transformers dependency: {exc}") from exc
+    except Exception as exc:
+        status = "degraded"
+        error = str(exc)
+
+    lock_path: Path | None = None
+    try:
+        lock_path = acquire_index_lock(index_dir)
+        has_index = idx_path.exists()
+        has_metadata = meta_path.exists()
+        has_manifest = manifest_path.exists()
+        rows: list[dict[str, Any]] | None = None
+
+        if has_index or has_metadata or has_manifest:
+            if not has_index or not has_metadata or not has_manifest:
+                if status == "ok":
+                    status = "degraded"
+                if not error:
+                    error = "conversation index artifacts incomplete; rebuild the FAISS conversation index"
+            else:
+                try:
+                    rows = read_metadata(meta_path)
+                    _np, faiss = load_vector_dependencies()
+                    index = faiss.read_index(str(idx_path))
+                    manifest_details = validate_index_manifest(
+                        read_manifest(manifest_path),
+                        requested_model_id=model_id,
+                        actual_dimension=int(index.d),
+                    )
+                    validate_artifact_counts(index, rows, manifest_details)
+                except Exception as exc:
+                    if status == "ok":
+                        status = "degraded"
+                    if not error:
+                        error = str(exc)
+        elif status == "ok":
+            status = "degraded"
+            error = "conversation index artifacts missing; build the FAISS conversation index"
+
+        if include_metadata:
+            if rows is None:
+                rows = read_metadata(meta_path)
+            metadata_details = {
+                "chunkCount": len(rows),
+                "hasIndex": has_index,
+                "hasMetadata": has_metadata,
+                "hasManifest": has_manifest,
+            }
+    except Exception as exc:
+        if status == "ok":
+            status = "degraded"
+        if not error:
+            error = str(exc)
+        if include_metadata:
+            try:
+                chunk_count = len(read_metadata(meta_path))
+            except Exception:
+                chunk_count = 0
+            metadata_details = {
+                "chunkCount": chunk_count,
+                "hasIndex": idx_path.exists(),
+                "hasMetadata": meta_path.exists(),
+                "hasManifest": manifest_path.exists(),
+            }
+    finally:
+        if lock_path is not None:
+            release_index_lock(lock_path)
+
+    response: dict[str, Any] = {"ok": True, "status": status}
+    if error:
+        response["error"] = error
+    if manifest_details is not None:
+        response["manifest"] = manifest_details
+    if metadata_details is not None:
+        response["metadata"] = metadata_details
+    return response
+
+
+def run_health(payload: dict[str, Any]) -> dict[str, Any]:
+    return build_health_response(payload)
+
+
+def run_inspect(payload: dict[str, Any]) -> dict[str, Any]:
+    return build_health_response(payload, include_metadata=True)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("command", choices=["upsert", "rebuild", "search", "health", "inspect"])
+    args = parser.parse_args()
+
+    try:
+        payload = read_payload()
+        if args.command == "upsert":
+            emit(run_upsert(payload))
+        elif args.command == "rebuild":
+            emit(run_rebuild(payload))
+        elif args.command == "search":
+            emit(run_search(payload))
+        elif args.command == "inspect":
+            emit(run_inspect(payload))
+        else:
+            emit(run_health(payload))
+        return 0
+    except (SidecarError, DependencyError) as exc:
+        emit({"ok": False, "error": str(exc)})
+        return 0
+    except Exception as exc:
+        print(f"faiss sidecar internal error: {exc}", file=sys.stderr)
+        emit({"ok": False, "error": "internal sidecar error"})
+        return 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())