mindnet/app/core/chunk_payload.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Modul: app/core/chunk_payload.py
Version: 2.2.1
Datum: 2025-11-07

Zweck
-----
Erzeugt Qdrant-Payloads für Chunks. Voll abwärtskompatibel zu v2.2.0 / v2.0.1.
Neu (2.2.1):
  • Stabilere Offsets (start/end) bei mehrfach vorkommenden Segmenten (inkrementelles Suchen + Fallback),
  • optionale Felder window_left_ctx_len und window_right_ctx_est zur Diagnose,
  • robustere Section-Pfadbehandlung.

Felder (unverändert beibehalten):
  note_id, chunk_id (Alias: id), chunk_index, seq, path,
  window (mit linkem Overlap), text (Kernsegment), start, end,
  overlap_left, overlap_right,
  type, title, tags, token_count?, section_title?, section_path?
"""
from __future__ import annotations

from typing import Any, Dict, Iterable, List, Optional, Tuple, Union

try:
    from app.core.chunk_config import get_sizes as _get_sizes
except Exception:
    def _get_sizes(_note_type: str):
        return {"overlap": (40, 60), "target": (250, 350), "max": 500}


def _get_attr_or_key(obj: Any, key: str, default=None):
    if obj is None:
        return default
    if isinstance(obj, dict):
        return obj.get(key, default)
    return getattr(obj, key, default)

def _as_window_text(chunk: Any) -> str:
    for k in ("window", "text", "content", "raw"):
        v = _get_attr_or_key(chunk, k, None)
        if isinstance(v, str) and v:
            return v
    return ""

def _to_int(x: Any, default: int = 0) -> int:
    try:
        return int(x)
    except Exception:
        return default

def _normalize_rel_path(p: str) -> str:
    p = (p or "").replace("\\", "/")
    while p.startswith("/"):
        p = p[1:]
    return p

def _dedupe_windows_to_segments(windows: List[str]) -> Tuple[List[str], List[int], str]:
    segments: List[str] = []
    overlaps_left: List[int] = []
    reconstructed = ""
    for w in windows:
        w = w or ""
        max_k = min(len(w), len(reconstructed))
        k = 0
        for cand in range(max_k, -1, -1):
            if reconstructed.endswith(w[:cand]):
                k = cand
                break
        seg = w[k:]
        segments.append(seg)
        overlaps_left.append(k)
        reconstructed += seg
    return segments, overlaps_left, reconstructed

def _overlap_len_suffix_prefix(a: str, b: str, max_probe: int = 4096) -> int:
    if not a or not b:
        return 0
    a1 = a[-max_probe:]
    b1 = b[:max_probe]
    n = min(len(a1), len(b1))
    for k in range(n, 0, -1):
        if a1[-k:] == b1[:k]:
            return k
    return 0

def make_chunk_payloads(
    frontmatter: Dict[str, Any],
    rel_path: str,
    chunks: Iterable[Union[Dict[str, Any], Any]],
    note_text: Optional[str] = None,
) -> List[Dict[str, Any]]:
    note_id = str(frontmatter.get("id") or "").strip()
    note_type = str(frontmatter.get("type", "")).lower()
    note_title = frontmatter.get("title", None)
    note_tags = frontmatter.get("tags", None)
    rel_path = _normalize_rel_path(rel_path)

    chunks_list = list(chunks)
    raw_windows: List[str] = []
    seqs: List[int] = []
    ids_in: List[Optional[str]] = []
    token_counts: List[Optional[int]] = []
    section_titles: List[Optional[str]] = []
    section_paths: List[Optional[str]] = []
    any_explicit_window = False

    for idx, c in enumerate(chunks_list):
        w = _get_attr_or_key(c, "window", None)
        if isinstance(w, str) and w:
            any_explicit_window = True
            raw_windows.append(w)
        else:
            raw_windows.append(_as_window_text(c))
        seqs.append(_to_int(_get_attr_or_key(c, "seq", _get_attr_or_key(c, "chunk_index", idx)), idx))
        cid = _get_attr_or_key(c, "chunk_id", _get_attr_or_key(c, "id", None))
        ids_in.append(str(cid) if isinstance(cid, str) and cid else None)
        tc = _get_attr_or_key(c, "token_count", None)
        token_counts.append(_to_int(tc, 0) if tc is not None else None)
        section_titles.append(_get_attr_or_key(c, "section_title", None))
        section_paths.append(_get_attr_or_key(c, "section_path", None))

    if any_explicit_window:
        segments, overlaps_left, recon = _dedupe_windows_to_segments(raw_windows)
        windows_final = raw_windows[:]
    else:
        segments = [w or "" for w in raw_windows]
        overlaps_left = []
        windows_final = []
        recon = ""
        try:
            overlap_low, overlap_high = tuple(_get_sizes(note_type).get("overlap", (40, 60)))
        except Exception:
            overlap_low, overlap_high = (40, 60)
        overlap_target = int(overlap_low)

        for i, seg in enumerate(segments):
            if i == 0:
                windows_final.append(seg)
                overlaps_left.append(0)
                recon += seg
            else:
                k = min(overlap_target, len(recon))
                left_ctx = recon[-k:] if k > 0 else ""
                windows_final.append(left_ctx + seg)
                overlaps_left.append(k)
                recon += seg

    overlaps_right: List[int] = []
    for i in range(len(windows_final)):
        if i + 1 < len(windows_final):
            ov = _overlap_len_suffix_prefix(windows_final[i], windows_final[i + 1], max_probe=4096)
        else:
            ov = 0
        overlaps_right.append(ov)

    starts: List[int] = [0] * len(segments)
    ends: List[int] = [0] * len(segments)
    if isinstance(note_text, str) and note_text:
        search_pos = 0
        for i, seg in enumerate(segments):
            if not seg:
                starts[i] = ends[i] = search_pos
                continue
            j = note_text.find(seg, search_pos)
            if j >= 0:
                starts[i] = j
                ends[i] = j + len(seg)
                search_pos = ends[i]
            else:
                # Fallback: naive fortlaufende Positionierung
                starts[i] = starts[i - 1] if i > 0 else 0
                ends[i] = starts[i] + len(seg)
                search_pos = ends[i]
    else:
        pos = 0
        for i, seg in enumerate(segments):
            starts[i] = pos
            pos += len(seg)
            ends[i] = pos

    payloads: List[Dict[str, Any]] = []
    for i, (win, seg) in enumerate(zip(windows_final, segments)):
        chunk_id = ids_in[i] or f"{note_id}#{i+1}"
        left_len = max(0, len(win) - len(seg))
        pl: Dict[str, Any] = {
            "note_id": note_id,
            "chunk_id": chunk_id,
            "id": chunk_id,
            "chunk_index": i,
            "seq": seqs[i],
            "path": rel_path,
            "window": win,
            "text": seg,
            "start": starts[i],
            "end": ends[i],
            "overlap_left": overlaps_left[i],
            "overlap_right": overlaps_right[i],
            "window_left_ctx_len": left_len,
        }
        if note_type:
            pl["type"] = note_type
        if note_title is not None:
            pl["title"] = note_title
        if note_tags is not None:
            pl["tags"] = note_tags
        if token_counts[i] is not None:
            pl["token_count"] = int(token_counts[i])
        if section_titles[i] is not None:
            pl["section_title"] = section_titles[i]
        if section_paths[i] is not None:
            sp = str(section_paths[i]).replace("\\", "/")
            pl["section_path"] = sp if sp else "/"
        payloads.append(pl)

    return payloads

if __name__ == "__main__":  # pragma: no cover
    fm = {"id": "demo", "title": "Demo", "type": "concept"}
    chunks = [
        {"id": "demo#1", "text": "Alpha Beta Gamma"},
        {"id": "demo#2", "text": "Gamma Delta"},
        {"id": "demo#3", "text": "Delta Epsilon Zeta"},
    ]
    pls = make_chunk_payloads(fm, "path/demo.md", chunks, note_text="Alpha Beta Gamma Delta Epsilon Zeta")
    from pprint import pprint
    pprint(pls)
    recon = "".join(p["text"] for p in pls)
    print("RECON:", recon)