Merge upstream main and resolve conflicts

.ai_update/changes.md

@@ -1,27 +1,125 @@
-# AI Update Log
+# AI Change Log
 
-## Scope
-- File: `search/main.py`
-- Purpose: fixed and extended retrieval/rerank pipeline according to TODO items.
+Дата создания: 2026-04-18
 
-## Done Changes
-- Switched base query selection to `question.search_text` with fallback to `question.text`.
-- Added support for `question.variants` as additional query variants in retrieval.
-- Added support for `question.hyde` as additional dense-only queries.
-- Added support for `question.keywords` as the primary source for sparse query text with fallback to current query variant.
-- Stopped losing retrieval candidates after rerank: rerank is applied to head (`RERANK_LIMIT`), tail candidates are preserved.
-- Added deduplication of retrieval points by Qdrant point id before rerank.
-- Implemented score aggregation by `message_id` (sum of chunk scores mapped to same message).
-- Limited final response to `top-50` message ids via `FINAL_TOP_K = 50`.
-- Final output message ids are now selected from aggregated scores (sorted by score desc, tie-break by message_id).
+## Как пользоваться
 
-## Notes
-- Earlier step introduced direct `message_ids` deduplication before response.
-- Current logic supersedes this by ranking and selecting unique `message_id` values from aggregated scores.
+Этот файл - рабочий журнал изменений для Codex.
 
-## Verification
-- Syntax check passed after each main change: `python -m py_compile search/main.py`.
+Перед новыми правками нужно прочитать этот файл и учитывать:
 
-## How To Use This Log
-- Treat this file as the source of truth for already completed `search/main.py` tasks.
-- On next tasks, read this file first to avoid duplicate edits.
+- что уже было изменено;
+- какие файлы трогались;
+- какие проверки запускались;
+- какие ограничения и договоренности есть по задаче.
+
+После каждой осмысленной правки нужно добавлять новую запись с:
+
+- кратким описанием изменения;
+- списком измененных файлов;
+- результатом проверок;
+- открытыми рисками или TODO, если они есть.
+
+## Договоренности
+
+- Не менять контракты `POST /index`, `POST /sparse_embedding`, `POST /search`.
+- По текущей задаче фокус держать на `index`, если пользователь не просит иначе.
+- Не трогать `docker-compose.yml` и инфраструктуру без отдельной просьбы.
+- Не откатывать чужие или пользовательские изменения.
+
+## Записи
+
+### 2026-04-18 - разрешены конфликты с upstream/main
+
+Что сделано:
+
+- Объединены записи журнала из ветки PR и `upstream/main`.
+- В `search/main.py` сохранены расширенные retrieval/rerank доработки из обеих веток.
+- Убраны конфликтные маркеры после merge `upstream/main`.
+- Сохранены безопасные лимиты retrieval и защита от падения rerank.
+
+Измененные файлы:
+
+- `.ai_update/changes.md`
+- `search/main.py`
+
+Проверки:
+
+- `python3 -m py_compile index/main.py search/main.py`
+- Прямой smoke `build_chunks` на `data/Go Nova.json`: 15 чанков, покрыто 25 из 25 сообщений.
+- Pure smoke для `search/main.py` helper-функций через stub-модули.
+
+Открытые риски:
+
+- Полный интеграционный прогон с настоящими Qdrant/dense/rerank зависит от внешних сервисов.
+
+### 2026-04-18 - index P1 and search retrieval/rerank
+
+Что сделано:
+
+- В `index/main.py` заменен символьный chunking на сборку чанков окнами сообщений.
+- Добавлен учет временного разрыва между сообщениями через `INDEX_TIME_GAP_SECONDS`.
+- Overlap теперь работает по границам сообщений через `INDEX_CHUNK_OVERLAP_MESSAGES`, а не по хвосту строки.
+- Сообщения рендерятся структурно: `author`, `time`, `thread`, `mentions`, флаги, `quote`, `forward`, `file`, `system_event`.
+- `file_snippets` парсятся как JSON; в индекс попадают имя файла, mime, url, владелец и дата создания.
+- `member_event` превращается в индексируемый системный текст.
+- `page_content`, `dense_content`, `sparse_content` разведены.
+- В `index/Dockerfile` старый `CHUNK_SIZE=10` заменен на реальные `INDEX_*` настройки chunking.
+- В `search/main.py` исправлен runtime-баг с неинициализированным `must_conditions`.
+- Основной query в search теперь берется из `question.search_text` с fallback на `question.text`.
+- `question.variants`, `question.hyde`, `question.keywords` и entities используются как дополнительные dense/sparse запросы.
+- Retrieval делает несколько prefetch-запросов и fusion через Qdrant.
+- Rerank больше не выбрасывает retrieval-хвост.
+- Retrieval points дедуплицируются по Qdrant point id.
+- Финальные `message_ids` агрегируются по score, дедуплицируются и ограничиваются `top-50`.
+
+Измененные файлы:
+
+- `index/main.py`
+- `index/Dockerfile`
+- `search/main.py`
+- `.ai_update/changes.md`
+
+Проверки:
+
+- `python3 -m py_compile index/main.py search/main.py`
+- Прямой smoke `build_chunks` на `data/Go Nova.json`.
+- Прямой smoke endpoint-функции `index(...)` на `data/Go Nova.json`.
+- Pure smoke для `search` helper-функций через stub-модули, потому в host env нет `qdrant_client` и `httpx`.
+
+Результаты проверки индекса:
+
+- Было 29 чанков, стало 15.
+- Покрытие сообщений на `data/Go Nova.json`: 25 из 25.
+- Системное сообщение с `member_event` больше не выпадает.
+- `file_snippets` с `IMG_8471.webp` попадает в `sparse_content`.
+- Quote и forward маркеры попадают в `dense_content`.
+- `page_content`, `dense_content`, `sparse_content` больше не одинаковые.
+
+Открытые риски:
+
+- Полный интеграционный прогон `search` с настоящими Qdrant/dense/rerank локально не выполнялся.
+- `date_range` фильтр включается только если установленный `qdrant_client` поддерживает `models.DatetimeRange`.
+- Полный docker build локально не запускался.
+
+### 2026-04-18 - создан журнал изменений
+
+Что сделано:
+
+- Создан файл `.ai_update/changes.md`.
+- Зафиксировано, что до этого код не менялся, была только разведка репозитория и ТЗ.
+
+Контекст по текущему состоянию:
+
+- `index/main.py` тогда использовал символьный chunking.
+- `render_message` брал только `message.text` и `parts[*].text`.
+- `page_content`, `dense_content`, `sparse_content` были одинаковые.
+- В примере `data/Go Nova.json` старый `build_chunks` покрывал 24 из 25 сообщений; системное сообщение с `member_event` выпадало из индекса.
+
+Измененные файлы:
+
+- `.ai_update/changes.md`
+
+Проверки:
+
+- Код не запускался, потому что создан только журнал.

.gitignore

@@ -214,3 +214,4 @@ __marimo__/
 
 # Streamlit
 .streamlit/secrets.toml
+.DS_Store

index/Dockerfile

@@ -9,7 +9,11 @@ COPY main.py .
 
 ENV HOST=0.0.0.0
 ENV PORT=8000
-ENV CHUNK_SIZE=10
+ENV INDEX_CHUNK_MAX_CHARS=2200
+ENV INDEX_MESSAGE_MAX_CHARS=1400
+ENV INDEX_TEXT_SECTION_MAX_CHARS=1000
+ENV INDEX_TIME_GAP_SECONDS=21600
+ENV INDEX_CHUNK_OVERLAP_MESSAGES=1
 ENV FASTEMBED_CACHE_PATH=/models/fastembed
 ENV HF_HOME=/models/huggingface
 

index/main.py

@@ -1,9 +1,12 @@
 import logging
 import os
+import json
+import re
+from dataclasses import dataclass
+from datetime import datetime, timezone
 from functools import lru_cache
 from typing import Any
 import asyncio
-import hashlib
 
 from fastapi import FastAPI, Request
 from fastapi.exceptions import RequestValidationError
@@ -88,31 +91,441 @@ app = FastAPI(title="Index Service", version="0.1.0")
 # Ваша внутренняя логика построения чанков. Можете делать всё, что посчитаете нужным.
 # Текущий код – минимальный пример
 
-CHUNK_SIZE = 512
-OVERLAP_SIZE = 256
+CHUNK_MAX_CHARS = int(os.getenv("INDEX_CHUNK_MAX_CHARS", "2200"))
+MESSAGE_MAX_CHARS = int(os.getenv("INDEX_MESSAGE_MAX_CHARS", "1400"))
+TEXT_SECTION_MAX_CHARS = int(os.getenv("INDEX_TEXT_SECTION_MAX_CHARS", "1000"))
+TIME_GAP_SECONDS = int(os.getenv("INDEX_TIME_GAP_SECONDS", str(6 * 60 * 60)))
+CHUNK_OVERLAP_MESSAGES = int(os.getenv("INDEX_CHUNK_OVERLAP_MESSAGES", "1"))
 SPARSE_MODEL_NAME = "Qdrant/bm25"
 FASTEMBED_CACHE_PATH = "/models/fastembed"
 
 # Важная переманная, которая позволяет вычислять sparse вектор в несколько ядер. Не рекомендуется изменять.
-UVICORN_WORKERS=8
+UVICORN_WORKERS = 8
 
-def render_message(message: Message) -> str:
-    text = ""
 
-    if message.text:
-        text += message.text
+@dataclass(frozen=True)
+class RenderedText:
+    page: str
+    dense: str
+    sparse: str
 
-    if message.parts:
-        parts_text: list[str] = []
-        for part in message.parts:
-            # parts различаются по своему типу, см. README.md
-            part_text = part.get("text")
-            if isinstance(part_text, str) and part_text:
-                parts_text.append(part_text)
-        if parts_text:
-            text += "\n".join(parts_text)
 
-    return text
+@dataclass(frozen=True)
+class RenderedUnit:
+    message_id: str
+    time: int
+    text: RenderedText
+
+
+@dataclass(frozen=True)
+class ChunkUnit:
+    unit: RenderedUnit
+    is_new: bool
+
+
+def clean_text(value: Any) -> str:
+    if not isinstance(value, str):
+        return ""
+
+    text = (
+        value.replace("\r\n", "\n")
+        .replace("\r", "\n")
+        .replace("\u200b", " ")
+        .replace("\xa0", " ")
+    )
+    lines = [re.sub(r"[ \t]+", " ", line).strip() for line in text.split("\n")]
+    result: list[str] = []
+    previous_blank = False
+
+    for line in lines:
+        if not line:
+            if result and not previous_blank:
+                result.append("")
+            previous_blank = True
+            continue
+
+        result.append(line)
+        previous_blank = False
+
+    return "\n".join(result).strip()
+
+
+def unique_preserve_order(values: list[str]) -> list[str]:
+    seen: set[str] = set()
+    result: list[str] = []
+
+    for value in values:
+        if value and value not in seen:
+            seen.add(value)
+            result.append(value)
+
+    return result
+
+
+def format_time(timestamp: int) -> str:
+    return datetime.fromtimestamp(timestamp, tz=timezone.utc).isoformat().replace("+00:00", "Z")
+
+
+def format_optional_time(value: Any) -> str:
+    if value is None or value == "":
+        return ""
+
+    try:
+        return format_time(int(value))
+    except (TypeError, ValueError, OSError, OverflowError):
+        return clean_text(str(value))
+
+
+def lexical_terms(value: str) -> str:
+    return clean_text(re.sub(r"[^0-9A-Za-zА-Яа-яЁё]+", " ", value))
+
+
+def split_long_text(text: str, max_chars: int) -> list[str]:
+    text = clean_text(text)
+    if not text:
+        return []
+    if len(text) <= max_chars:
+        return [text]
+
+    paragraphs = [item.strip() for item in re.split(r"\n{2,}", text) if item.strip()]
+    pieces: list[str] = []
+    current = ""
+
+    def append_current() -> None:
+        nonlocal current
+        if current:
+            pieces.append(current)
+            current = ""
+
+    def split_oversized(paragraph: str) -> list[str]:
+        words = paragraph.split()
+        result: list[str] = []
+        part = ""
+
+        for word in words:
+            if not part:
+                part = word
+                continue
+
+            if len(part) + 1 + len(word) <= max_chars:
+                part += " " + word
+            else:
+                result.append(part)
+                part = word
+
+        if part:
+            result.append(part)
+
+        return result
+
+    for paragraph in paragraphs:
+        candidates = [paragraph] if len(paragraph) <= max_chars else split_oversized(paragraph)
+        for candidate in candidates:
+            separator = "\n\n" if current else ""
+            if current and len(current) + len(separator) + len(candidate) > max_chars:
+                append_current()
+
+            current = candidate if not current else current + separator + candidate
+
+    append_current()
+    return pieces
+
+
+def combine_texts(items: list[RenderedText], separator: str = "\n") -> RenderedText:
+    return RenderedText(
+        page=separator.join(item.page for item in items if item.page).strip(),
+        dense=separator.join(item.dense for item in items if item.dense).strip(),
+        sparse=separator.join(item.sparse for item in items if item.sparse).strip(),
+    )
+
+
+def rendered_length(text: RenderedText) -> int:
+    return max(len(text.page), len(text.dense), len(text.sparse))
+
+
+def message_header(message: Message) -> RenderedText:
+    timestamp = format_time(message.time)
+    mentions = unique_preserve_order(message.mentions or [])
+
+    page_lines = [
+        f"author: {message.sender_id}",
+        f"time: {timestamp}",
+    ]
+    dense_lines = [
+        f"author: {message.sender_id}",
+        f"message_time: {timestamp}",
+    ]
+    sparse_terms = [
+        "author",
+        message.sender_id,
+        lexical_terms(message.sender_id),
+        timestamp,
+    ]
+
+    if message.thread_sn:
+        page_lines.append(f"thread: {message.thread_sn}")
+        dense_lines.append(f"thread: {message.thread_sn}")
+        sparse_terms.extend(["thread", message.thread_sn, lexical_terms(message.thread_sn)])
+
+    if mentions:
+        mentions_text = ", ".join(mentions)
+        page_lines.append(f"mentions: {mentions_text}")
+        dense_lines.append(f"mentions: {mentions_text}")
+        sparse_terms.extend(["mentions", *mentions, *(lexical_terms(item) for item in mentions)])
+
+    flags = []
+    if message.is_system:
+        flags.append("system")
+    if message.is_forward:
+        flags.append("forward")
+    if message.is_quote:
+        flags.append("quote")
+    if message.is_hidden:
+        flags.append("hidden")
+
+    if flags:
+        flags_text = ", ".join(flags)
+        dense_lines.append(f"message_flags: {flags_text}")
+        sparse_terms.extend(flags)
+
+    return RenderedText(
+        page="\n".join(page_lines),
+        dense="\n".join(dense_lines),
+        sparse=" ".join(term for term in sparse_terms if term),
+    )
+
+
+def render_plain_sections(text: str, label: str = "text") -> list[RenderedText]:
+    sections: list[RenderedText] = []
+    pieces = split_long_text(text, TEXT_SECTION_MAX_CHARS)
+
+    for index, piece in enumerate(pieces):
+        suffix = f" part {index + 1}/{len(pieces)}" if len(pieces) > 1 else ""
+        sections.append(
+            RenderedText(
+                page=piece,
+                dense=f"{label}{suffix}: {piece}",
+                sparse=f"{label} {piece}",
+            )
+        )
+
+    return sections
+
+
+def render_part_sections(part: dict[str, Any]) -> list[RenderedText]:
+    text = clean_text(part.get("text"))
+    if not text:
+        return []
+
+    media_type = clean_text(part.get("mediaType") or part.get("type") or "text").lower()
+    source = clean_text(part.get("sn"))
+    part_time = part.get("time")
+    source_bits = []
+    if source:
+        source_bits.append(f"source: {source}")
+    formatted_part_time = format_optional_time(part_time)
+    if formatted_part_time:
+        source_bits.append(f"source_time: {formatted_part_time}")
+    source_text = ", ".join(source_bits)
+
+    if media_type == "quote":
+        label = f"quote from {source}" if source else "quote"
+        dense_label = f"quote; {source_text}" if source_text else "quote"
+        sparse_prefix = f"quote цитата {source} {lexical_terms(source)}"
+    elif media_type == "forward":
+        label = f"forwarded from {source}" if source else "forwarded"
+        dense_label = f"forwarded_message; {source_text}" if source_text else "forwarded_message"
+        sparse_prefix = f"forward forwarded_message пересланное {source} {lexical_terms(source)}"
+    else:
+        label = "text"
+        dense_label = "text"
+        sparse_prefix = "text"
+
+    sections: list[RenderedText] = []
+    pieces = split_long_text(text, TEXT_SECTION_MAX_CHARS)
+
+    for index, piece in enumerate(pieces):
+        suffix = f" part {index + 1}/{len(pieces)}" if len(pieces) > 1 else ""
+        page_prefix = f"{label}{suffix}:"
+        sections.append(
+            RenderedText(
+                page=f"{page_prefix}\n{piece}" if media_type in {"quote", "forward"} else piece,
+                dense=f"{dense_label}{suffix}: {piece}",
+                sparse=f"{sparse_prefix} {piece}",
+            )
+        )
+
+    return sections
+
+
+def render_file_sections(raw_snippets: str) -> list[RenderedText]:
+    raw_snippets = clean_text(raw_snippets)
+    if not raw_snippets:
+        return []
+
+    try:
+        parsed = json.loads(raw_snippets)
+    except json.JSONDecodeError:
+        return [
+            RenderedText(
+                page=f"file_snippet: {raw_snippets}",
+                dense=f"file_snippet: {raw_snippets}",
+                sparse=f"file file_snippet {raw_snippets}",
+            )
+        ]
+
+    snippets = parsed if isinstance(parsed, list) else [parsed]
+    sections: list[RenderedText] = []
+
+    for snippet in snippets:
+        if not isinstance(snippet, dict):
+            text = clean_text(str(snippet))
+            sections.append(RenderedText(page=f"file: {text}", dense=f"file: {text}", sparse=f"file {text}"))
+            continue
+
+        name = clean_text(snippet.get("name"))
+        mime = clean_text(snippet.get("mime"))
+        url = clean_text(snippet.get("original_url") or snippet.get("url"))
+        owner = clean_text(snippet.get("uid"))
+        created = clean_text(snippet.get("date_create"))
+        file_bits = [
+            f"name: {name}" if name else "",
+            f"mime: {mime}" if mime else "",
+            f"url: {url}" if url else "",
+            f"owner: {owner}" if owner else "",
+            f"created: {created}" if created else "",
+        ]
+        file_text = ", ".join(bit for bit in file_bits if bit)
+        sparse_terms = " ".join(
+            term
+            for term in [
+                "file",
+                "attachment",
+                "document",
+                name,
+                lexical_terms(name),
+                mime,
+                url,
+                owner,
+                lexical_terms(owner),
+                created,
+            ]
+            if term
+        )
+        sections.append(
+            RenderedText(
+                page=f"file: {file_text}",
+                dense=f"file: {file_text}",
+                sparse=sparse_terms,
+            )
+        )
+
+    return sections
+
+
+def render_member_event(message: Message) -> list[RenderedText]:
+    event = message.member_event
+    if not event:
+        return []
+
+    event_type = clean_text(event.get("type") or "member_event")
+    members_raw = event.get("members")
+    members = [clean_text(item) for item in members_raw] if isinstance(members_raw, list) else []
+    members = unique_preserve_order([item for item in members if item])
+
+    if members:
+        members_text = ", ".join(members)
+    else:
+        members_text = " ".join(clean_text(str(value)) for value in event.values() if value)
+
+    page = f"system_event: {event_type}; actor: {message.sender_id}; members: {members_text}"
+    dense = (
+        f"system_event: {event_type}; action: add or update chat members; "
+        f"actor: {message.sender_id}; members: {members_text}"
+    )
+    sparse = " ".join(
+        term
+        for term in [
+            "system_event",
+            "member_event",
+            event_type,
+            "addMembers",
+            "добавление участников",
+            message.sender_id,
+            lexical_terms(message.sender_id),
+            members_text,
+            lexical_terms(members_text),
+        ]
+        if term
+    )
+
+    return [RenderedText(page=page, dense=dense, sparse=sparse)]
+
+
+def render_message_sections(message: Message) -> list[RenderedText]:
+    sections: list[RenderedText] = []
+    sections.extend(render_plain_sections(message.text, "message_text"))
+
+    for part in message.parts or []:
+        if isinstance(part, dict):
+            sections.extend(render_part_sections(part))
+
+    sections.extend(render_file_sections(message.file_snippets))
+    sections.extend(render_member_event(message))
+
+    return sections
+
+
+def render_message_units(message: Message) -> list[RenderedUnit]:
+    sections = render_message_sections(message)
+    if not sections:
+        return []
+
+    header = message_header(message)
+    units: list[RenderedUnit] = []
+    current: list[RenderedText] = []
+
+    def flush() -> None:
+        nonlocal current
+        if not current:
+            return
+
+        text = combine_texts([header, *current])
+        units.append(RenderedUnit(message_id=message.id, time=message.time, text=text))
+        current = []
+
+    for section in sections:
+        candidate = combine_texts([header, *current, section])
+        if current and rendered_length(candidate) > MESSAGE_MAX_CHARS:
+            flush()
+
+        current.append(section)
+
+    flush()
+    return units
+
+
+def chunk_text(items: list[ChunkUnit]) -> RenderedText:
+    return RenderedText(
+        page="\n\n".join(item.unit.text.page for item in items if item.unit.text.page).strip(),
+        dense="\n\n".join(item.unit.text.dense for item in items if item.unit.text.dense).strip(),
+        sparse="\n\n".join(item.unit.text.sparse for item in items if item.unit.text.sparse).strip(),
+    )
+
+
+def chunk_length(items: list[ChunkUnit]) -> int:
+    return rendered_length(chunk_text(items))
+
+
+def trim_context(context: list[RenderedUnit], unit: RenderedUnit) -> list[RenderedUnit]:
+    result = context[-CHUNK_OVERLAP_MESSAGES:] if CHUNK_OVERLAP_MESSAGES > 0 else []
+    items = [ChunkUnit(item, False) for item in result] + [ChunkUnit(unit, True)]
+
+    while result and chunk_length(items) > CHUNK_MAX_CHARS:
+        result = result[1:]
+        items = [ChunkUnit(item, False) for item in result] + [ChunkUnit(unit, True)]
+
+    return result
 
 
 def build_chunks(
@@ -121,71 +534,78 @@ def build_chunks(
 ) -> list[IndexAPIItem]:
     result: list[IndexAPIItem] = []
 
-    def build_text_and_ranges(messages: list[Message]) -> tuple[str, list[tuple[int, int, str]]]:
-        text_parts: list[str] = []
-        message_ranges: list[tuple[int, int, str]] = []
-        position = 0
-
-        for index, message in enumerate(messages):
-            text = render_message(message)
-            if not text:
-                continue
-
-            if index > 0 and text_parts:
-                text_parts.append("\n")
-                position += 1
-
-            start = position
-            text_parts.append(text)
-            position += len(text)
-            message_ranges.append((start, position, message.id))
-
-        return "".join(text_parts), message_ranges
-
-    def slice_tail(
-        text: str,
-        tail_size: int,
-    ) -> str:
-        if tail_size <= 0:
-            return ""
-
-        tail_start = max(0, len(text) - tail_size)
-        return text[tail_start:]
-
-    overlap_text, overlap_message_ranges = build_text_and_ranges(overlap_messages)
-    previous_chunk_text = slice_tail(overlap_text, OVERLAP_SIZE)
-
-    new_text, new_message_ranges = build_text_and_ranges(new_messages)
-
-    for start in range(0, len(new_text), CHUNK_SIZE):
-        chunk_body = new_text[start : start + CHUNK_SIZE]
-        if not chunk_body:
-            continue
-
-        chunk_body_ranges = [
-            (
-                max(message_start, start) - start,
-                min(message_end, start + len(chunk_body)) - start,
-                message_id,
-            )
-            for message_start, message_end, message_id in new_message_ranges
-            if message_end > start and message_start < start + len(chunk_body)
+    overlap_units = [
+        unit
+        for message in overlap_messages
+        for unit in render_message_units(message)
     ]
-        chunk_overlap = previous_chunk_text
-        chunk_text = chunk_overlap
-        if chunk_text and chunk_body:
-            chunk_text += "\n"
-        chunk_text += chunk_body
+    new_units = [
+        unit
+        for message in new_messages
+        for unit in render_message_units(message)
+    ]
+    current: list[ChunkUnit] = []
+    last_new_time: int | None = None
 
+    def flush_current() -> None:
+        nonlocal current
+        if not current:
+            return
+
+        message_ids = unique_preserve_order(
+            [item.unit.message_id for item in current if item.is_new]
+        )
+        if not message_ids:
+            current = []
+            return
+
+        text = chunk_text(current)
         result.append(
             IndexAPIItem(
-                page_content=chunk_text,
-                dense_content=chunk_text,
-                sparse_content=chunk_text,
-                message_ids=[message_id for _, _, message_id in chunk_body_ranges],
+                page_content=text.page,
+                dense_content=text.dense,
+                sparse_content=text.sparse,
+                message_ids=message_ids,
             )
         )
-        previous_chunk_text = slice_tail(chunk_text, OVERLAP_SIZE)
+        current = []
+
+    def request_overlap_context(unit: RenderedUnit) -> list[RenderedUnit]:
+        close_units = [
+            item
+            for item in overlap_units
+            if abs(unit.time - item.time) <= TIME_GAP_SECONDS
+        ]
+        return trim_context(close_units, unit)
+
+    for unit in new_units:
+        if not current:
+            context = request_overlap_context(unit)
+            current = [ChunkUnit(item, False) for item in context]
+            current.append(ChunkUnit(unit, True))
+            last_new_time = unit.time
+            continue
+
+        gap = abs(unit.time - last_new_time) if last_new_time is not None else 0
+        candidate = [*current, ChunkUnit(unit, True)]
+        should_split = gap > TIME_GAP_SECONDS or chunk_length(candidate) > CHUNK_MAX_CHARS
+
+        if should_split:
+            previous_new_units = [item.unit for item in current if item.is_new]
+            context = (
+                trim_context(previous_new_units, unit)
+                if gap <= TIME_GAP_SECONDS
+                else request_overlap_context(unit)
+            )
+            flush_current()
+            current = [ChunkUnit(item, False) for item in context]
+            current.append(ChunkUnit(unit, True))
+        else:
+            current.append(ChunkUnit(unit, True))
+
+        last_new_time = unit.time
+
+    flush_current()
 
     return result
 

search/main.py

@@ -171,11 +171,14 @@ app = FastAPI(title="Search Service", version="0.1.0", lifespan=lifespan)
 # Внутри шаблона dense и rerank берутся из внешних HTTP endpoint'ов,
 # которые предоставляет проверяющая система.
 # Текущий код ниже — минимальный пример search pipeline.
-DENSE_PREFETCH_K = 10
-SPRASE_PREFETCH_K = 30
-RETRIEVE_K = 20
-RERANK_LIMIT = 10
+DENSE_PREFETCH_K = 30
+SPARSE_PREFETCH_K = 40
+RETRIEVE_K = 80
+RERANK_LIMIT = 20
 FINAL_TOP_K = 50
+MAX_DENSE_QUERIES = 4
+MAX_SPARSE_QUERIES = 3
+
 
 async def embed_dense(client: httpx.AsyncClient, text: str) -> list[float]:
     # Dense endpoint ожидает OpenAI-compatible body с input как списком строк.
@@ -207,108 +210,116 @@ async def embed_sparse(text: str) -> SparseVector:
         values=[float(value) for value in item.values.tolist()],
     )
 
-# ПЕРЕПИСАТЬ
 
-async def qdrant_search(
-    client: AsyncQdrantClient,
-    dense_vector: list[float],
-    sparse_vector: SparseVector,
-    question_data: Question
-) -> Any | None:
-    must_conditions: []
+def unique_non_empty(values: list[str | None]) -> list[str]:
+    result: list[str] = []
+    seen: set[str] = set()
 
-    # Фильтр по диапазону дат (поле metadata.start в Qdrant) [cite: 147, 148, 175]
-    if question_data.date_range:
+    for value in values:
+        text = (value or "").strip()
+        if text and text not in seen:
+            seen.add(text)
+            result.append(text)
+
+    return result
+
+
+def question_entity_terms(question: Question) -> list[str]:
+    entities = question.entities
+    if entities is None:
+        return []
+
+    values: list[str | None] = []
+    values.extend(entities.people or [])
+    values.extend(entities.emails or [])
+    values.extend(entities.documents or [])
+    values.extend(entities.names or [])
+    values.extend(entities.links or [])
+    return unique_non_empty(values)
+
+
+def build_query_texts(question: Question) -> tuple[str, list[str], list[str]]:
+    primary_query = (question.search_text or question.text).strip()
+    dense_queries = unique_non_empty(
+        [
+            primary_query,
+            *(question.variants or []),
+            *(question.hyde or []),
+        ]
+    )[:MAX_DENSE_QUERIES]
+
+    keyword_query = " ".join(question.keywords or []).strip()
+    entity_query = " ".join(question_entity_terms(question)).strip()
+    sparse_queries = unique_non_empty(
+        [
+            keyword_query,
+            entity_query,
+            primary_query,
+            *(question.variants or []),
+        ]
+    )[:MAX_SPARSE_QUERIES]
+
+    return primary_query, dense_queries, sparse_queries
+
+
+def build_search_filter(question: Question) -> models.Filter | None:
+    must_conditions: list[Any] = []
+
+    if question.date_range and hasattr(models, "DatetimeRange"):
         must_conditions.append(
             models.FieldCondition(
                 key="metadata.start",
-                range=models.Range(
-                    gte=question_data.date_range.from_,
-                    lte=question_data.date_range.to_
-                )
-            )
-        )
-
-    # Фильтр по автору вопроса (поле metadata.participants) [cite: 161, 163]
-    if question_data.asker:
-        must_conditions.append(
-            models.FieldCondition(
-                key="metadata.participants",
-                match=models.MatchValue(value=question_data.asker)
+                range=models.DatetimeRange(
+                    gte=question.date_range.from_,
+                    lte=question.date_range.to,
+                ),
             )
         )
 
-    # Создаем итоговый объект фильтра, если есть условия
-    search_filter = models.Filter(must=must_conditions) if must_conditions else None
+    return models.Filter(must=must_conditions) if must_conditions else None
 
-    response = await client.query_points(
-        collection_name=QDRANT_COLLECTION_NAME,
-        prefetch=[
+
+async def qdrant_search(
+    client: AsyncQdrantClient,
+    dense_vectors: list[list[float]],
+    sparse_vectors: list[SparseVector],
+    question_data: Question,
+) -> Any | None:
+    search_filter = build_search_filter(question_data)
+    prefetch: list[models.Prefetch] = []
+
+    for dense_vector in dense_vectors:
+        prefetch.append(
             models.Prefetch(
                 query=dense_vector,
                 using=QDRANT_DENSE_VECTOR_NAME,
                 limit=DENSE_PREFETCH_K,
                 filter=search_filter,
-            ),
+            )
+        )
+
+    for sparse_vector in sparse_vectors:
+        if not sparse_vector.indices:
+            continue
+
+        prefetch.append(
             models.Prefetch(
                 query=models.SparseVector(
                     indices=sparse_vector.indices,
                     values=sparse_vector.values,
                 ),
                 using=QDRANT_SPARSE_VECTOR_NAME,
-                limit=SPRASE_PREFETCH_K,
+                limit=SPARSE_PREFETCH_K,
                 filter=search_filter,
-            ),
-        ],
-        query=models.FusionQuery(fusion=models.Fusion.RRF),
-        limit=RETRIEVE_K,
-        with_payload=True,
+            )
         )
 
-    if not response.points:
+    if not prefetch:
         return None
 
-    return response.points
-
-
-async def qdrant_search_dense_only(
-    client: AsyncQdrantClient,
-    dense_vector: list[float],
-    question_data: Question,
-) -> Any | None:
-    must_conditions: []
-
-    if question_data.date_range:
-        must_conditions.append(
-            models.FieldCondition(
-                key="metadata.start",
-                range=models.Range(
-                    gte=question_data.date_range.from_,
-                    lte=question_data.date_range.to_,
-                ),
-            )
-        )
-
-    if question_data.asker:
-        must_conditions.append(
-            models.FieldCondition(
-                key="metadata.participants",
-                match=models.MatchValue(value=question_data.asker),
-            )
-        )
-
-    search_filter = models.Filter(must=must_conditions) if must_conditions else None
-
     response = await client.query_points(
         collection_name=QDRANT_COLLECTION_NAME,
-        prefetch=[
-            models.Prefetch(
-                query=dense_vector,
-                using=QDRANT_DENSE_VECTOR_NAME,
-                limit=DENSE_PREFETCH_K,
-                filter=search_filter,
-            ),
-        ],
+        prefetch=prefetch,
         query=models.FusionQuery(fusion=models.Fusion.RRF),
         limit=RETRIEVE_K,
         with_payload=True,
@@ -320,73 +331,17 @@ async def qdrant_search_dense_only(
     return response.points
 
 
-def collect_query_variants(question: Question) -> list[str]:
-    variants: list[str] = []
-    seen: set[str] = set()
-
-    def add_query(text: str | None) -> None:
-        if text is None:
-            return
-        normalized = text.strip()
-        if not normalized:
-            return
-        if normalized in seen:
-            return
-        seen.add(normalized)
-        variants.append(normalized)
-
-    add_query(question.search_text)
-    add_query(question.text)
-
-    for variant in question.variants or []:
-        add_query(variant)
-
-    return variants
-
-
-def collect_hyde_queries(question: Question, base_queries: list[str]) -> list[str]:
-    hyde_queries: list[str] = []
-    seen: set[str] = set(base_queries)
-
-    for hyde_query in question.hyde or []:
-        normalized = hyde_query.strip()
-        if not normalized:
-            continue
-        if normalized in seen:
-            continue
-        seen.add(normalized)
-        hyde_queries.append(normalized)
-
-    return hyde_queries
-
-
-def build_sparse_query_text(question: Question, fallback_query: str) -> str:
-    keywords: list[str] = []
-    seen: set[str] = set()
-
-    for keyword in question.keywords or []:
-        normalized = keyword.strip()
-        if not normalized:
-            continue
-        if normalized in seen:
-            continue
-        seen.add(normalized)
-        keywords.append(normalized)
-
-    if keywords:
-        return " ".join(keywords)
-
-    return fallback_query
-
-
 def deduplicate_points(points: list[Any]) -> list[Any]:
     unique_points: list[Any] = []
     seen_ids: set[str] = set()
 
     for point in points:
-        point_id = str(getattr(point, "id", ""))
-        if not point_id:
+        point_id = getattr(point, "id", None)
+        if point_id is None:
+            unique_points.append(point)
             continue
+
+        point_id = str(point_id)
         if point_id in seen_ids:
             continue
         seen_ids.add(point_id)
@@ -444,44 +399,64 @@ async def rerank_points(
 ) -> list[tuple[Any, float]]:
     rerank_candidates = points[:RERANK_LIMIT]
     tail_candidates = points[RERANK_LIMIT:]
-    rerank_targets = [point.payload.get("page_content") for point in rerank_candidates]
+    rerank_targets = [
+        str((point.payload or {}).get("page_content") or "")
+        for point in rerank_candidates
+    ]
+
+    try:
         scores = await get_rerank_scores(client, query, rerank_targets)
+    except Exception:
+        logger.exception("Rerank failed, returning retrieval order")
+        return [(point, extract_point_score(point)) for point in points]
+
+    if len(scores) != len(rerank_candidates):
+        logger.warning(
+            "Rerank returned %d scores for %d candidates",
+            len(scores),
+            len(rerank_candidates),
+        )
+        return [(point, extract_point_score(point)) for point in points]
 
     reranked_candidates = [
         (point, float(score))
         for score, point in sorted(
-            zip(scores, rerank_candidates, strict=True),
+            zip(scores, rerank_candidates),
             key=lambda item: item[0],
             reverse=True,
         )
     ]
     tail_with_scores = [
         (point, extract_point_score(point))
-        for _, point in sorted(
-            [(extract_point_score(point), point) for point in tail_candidates],
-            key=lambda item: item[0],
-            reverse=True,
-        )
+        for point in tail_candidates
     ]
 
     return reranked_candidates + tail_with_scores
 
 
-def aggregate_message_scores(scored_points: list[tuple[Any, float]]) -> dict[str, float]:
+def aggregate_message_scores(
+    scored_points: list[tuple[Any, float]],
+) -> tuple[dict[str, float], dict[str, int]]:
     aggregated_scores: dict[str, float] = {}
+    first_seen_rank: dict[str, int] = {}
 
-    for point, point_score in scored_points:
+    for rank, (point, point_score) in enumerate(scored_points):
         point_message_ids = set(extract_message_ids(point))
         for message_id in point_message_ids:
             aggregated_scores[message_id] = aggregated_scores.get(message_id, 0.0) + point_score
+            first_seen_rank.setdefault(message_id, rank)
 
-    return aggregated_scores
+    return aggregated_scores, first_seen_rank
 
 
-def select_top_message_ids(aggregated_scores: dict[str, float], limit: int) -> list[str]:
+def select_top_message_ids(
+    aggregated_scores: dict[str, float],
+    first_seen_rank: dict[str, int],
+    limit: int,
+) -> list[str]:
     sorted_items = sorted(
         aggregated_scores.items(),
-        key=lambda item: (-item[1], item[0]),
+        key=lambda item: (-item[1], first_seen_rank[item[0]]),
     )
     return [message_id for message_id, _ in sorted_items[:limit]]
 
@@ -494,37 +469,23 @@ async def health() -> dict[str, str]:
 
 @app.post("/search", response_model=SearchAPIResponse)
 async def search(payload: SearchAPIRequest) -> SearchAPIResponse:
-    queries = collect_query_variants(payload.question)
-    if not queries:
+    query, dense_queries, sparse_queries = build_query_texts(payload.question)
+    if not query:
         raise HTTPException(status_code=400, detail="question.search_text or question.text is required")
 
-    hyde_queries = collect_hyde_queries(payload.question, queries)
-    query = queries[0]
     client: httpx.AsyncClient = app.state.http
     qdrant: AsyncQdrantClient = app.state.qdrant
 
-    all_points: list[Any] = []
-    for query_variant in queries:
-        dense_vector = await embed_dense(client, query_variant)
-        sparse_query_text = build_sparse_query_text(payload.question, query_variant)
-        sparse_vector = await embed_sparse(sparse_query_text)
-        points = await qdrant_search(qdrant, dense_vector, sparse_vector, payload.question)
-        if points:
-            all_points.extend(list(points))
+    dense_vectors = [await embed_dense(client, item) for item in dense_queries]
+    sparse_vectors = [await embed_sparse(item) for item in sparse_queries]
+    best_points = await qdrant_search(qdrant, dense_vectors, sparse_vectors, payload.question)
 
-    for hyde_query in hyde_queries:
-        hyde_dense_vector = await embed_dense(client, hyde_query)
-        hyde_points = await qdrant_search_dense_only(qdrant, hyde_dense_vector, payload.question)
-        if hyde_points:
-            all_points.extend(list(hyde_points))
-
-    best_points = deduplicate_points(all_points)
     if not best_points:
         return SearchAPIResponse(results=[])
 
-    scored_points = await rerank_points(client, query, list(best_points))
-    aggregated_scores = aggregate_message_scores(scored_points)
-    message_ids = select_top_message_ids(aggregated_scores, FINAL_TOP_K)
+    scored_points = await rerank_points(client, query, deduplicate_points(list(best_points)))
+    aggregated_scores, first_seen_rank = aggregate_message_scores(scored_points)
+    message_ids = select_top_message_ids(aggregated_scores, first_seen_rank, FINAL_TOP_K)
 
     return SearchAPIResponse(
         results=[SearchAPIItem(message_ids=message_ids)]