fix(ai,reader): robust sentence segmentation without abbreviation cutoffs & context-aware literary translation

- Implement robustSentenceSplit across reader.js and backend (miner.py, cefr.py) to prevent false cutoffs on honorifics (Mr., Mrs., Dr.), single-letter initials (John F. Kennedy), decimals, acronyms (U.S., U.K.), and abbreviations (No. 4, etc.)
- Upgrade translation prompt to professional literary Russian, eliminating literal word-for-word translation and preserving titles/honorifics
- Inject surrounding paragraph context into translation API to avoid out-of-context translations
- Inject offline dictionary hints into LLM prompt for accurate contextual word meaning
- Prefer qwen2.5:3b for higher translation fluency and fidelity
- Purge truncated cache entries from sentence_translations and restore Card 30 (briefcase) with authentic full sentence
- Bump Service Worker cache to notes-v6.0
- Add unit tests for robust sentence splitting, prompt formatting, and translation API

app/ai.py

@@ -1,46 +1,81 @@
 import json
 import logging
+import re
 from typing import Optional, Dict, Any
 import httpx
 from app import config
-from app.database import get_cached_sentence_translation, save_sentence_translation
+from app.database import get_db, get_cached_sentence_translation, save_sentence_translation
 
 logger = logging.getLogger(__name__)
 
 
-def build_translation_prompt(sentence: str, target_word: str = "") -> str:
+def get_dictionary_hint(term: str) -> str:
+    """Retrieves verified offline dictionary translation hints for term."""
+    clean = term.strip()
+    if not clean:
+        return ""
+    try:
+        with get_db() as conn:
+            row = conn.execute(
+                "SELECT translation FROM offline_dictionary WHERE term = ? COLLATE NOCASE",
+                (clean,),
+            ).fetchone()
+            if row and row["translation"]:
+                return row["translation"].strip()
+    except Exception:
+        pass
+    return ""
+
+
+def build_translation_prompt(
+    sentence: str,
+    target_word: str = "",
+    context: str = "",
+    dict_hint: str = "",
+) -> str:
     clean_sentence = sentence.strip()
     clean_word = target_word.strip()
+    clean_context = context.strip()
 
+    context_section = ""
+    if clean_context and clean_context != clean_sentence:
+        short_ctx = clean_context[:500].strip()
+        context_section = f'Контекст абзаца:\n"{short_ctx}"\n\n'
+
+    target_section = ""
     if clean_word:
+        target_section = f'Целевое слово/выражение: "{clean_word}"\n'
+        if dict_hint:
+            target_section += f'Возможные словарные значения: {dict_hint}\n'
+        target_section += "\n"
+
     return (
-            "You are a professional English-to-Russian translator.\n"
+        "Ты — профессиональный литературный переводчик художественной и научно-популярной литературы на русский язык.\n"
-            "Translate the following English sentence into clear, natural, and idiomatic Russian.\n"
+        "Твоя задача — перевести указанное английское предложение на красивый, естественный и грамотный русский литературный язык, идеально вписывающийся в повествование книги.\n\n"
-            f"Also provide the precise in-context meaning and translation of the target word/phrase '{clean_word}'.\n\n"
+        f"{context_section}"
-            f'Sentence: "{clean_sentence}"\n'
+        f'Предложение для перевода:\n"{clean_sentence}"\n\n'
-            f'Target word: "{clean_word}"\n\n'
+        f"{target_section}"
-            "Respond ONLY with a valid JSON object in this format:\n"
+        "Обязательные правила перевода:\n"
+        "1. Перевод должен быть живым и художественным (НЕ дословным, без топорного калькирования).\n"
+        "2. Переводи всё предложение ПОЛНОСТЬЮ от начала до конца, не обрывая на точках, обращениях или знаках препинания.\n"
+        "3. Обязательно сохраняй обращения и титулы: Mr. -> мистер, Mrs. -> миссис, Dr. -> доктор, Prof. -> профессор, St. -> святой/Сент-. Не теряй их!\n"
+        "4. В поле \"word_meaning\" укажи только точный перевод выделенного слова на русский язык в контексте данного предложения (выбрав самый подходящий вариант из словарных подсказок).\n\n"
+        "Ответь ИСКЛЮЧИТЕЛЬНО в формате JSON:\n"
         "{\n"
-            '  "sentence_ru": "перевод предложения на русский язык",\n'
+        '  "sentence_ru": "полный литературный перевод всего предложения",\n'
-            '  "word_meaning": "значение слова в этом контексте"\n'
+        f'  "word_meaning": "{"русский перевод слова в данном контексте" if clean_word else ""}"\n'
-            "}"
-        )
-    else:
-        return (
-            "You are a professional English-to-Russian translator.\n"
-            "Translate the following English sentence into clear, natural, and idiomatic Russian.\n\n"
-            f'Sentence: "{clean_sentence}"\n\n'
-            "Respond ONLY with a valid JSON object in this format:\n"
-            "{\n"
-            '  "sentence_ru": "перевод предложения на русский язык",\n'
-            '  "word_meaning": ""\n'
         "}"
     )
 
 
-async def translate_sentence_ai(sentence: str, target_word: str = "") -> Dict[str, Any]:
+async def translate_sentence_ai(
+    sentence: str,
+    target_word: str = "",
+    context: str = ""
+) -> Dict[str, Any]:
     """
     Translates a sentence using local AI (Ollama), with SQLite instant-cache lookup.
+    Incorporates paragraph context and offline dictionary hints for maximum naturalness.
     Returns:
     {
         "original_sentence": "...",
@@ -53,6 +88,7 @@ async def translate_sentence_ai(sentence: str, target_word: str = "") -> Dict[st
     """
     clean_sentence = sentence.strip()
     clean_word = target_word.strip()
+    clean_context = context.strip()
 
     if not clean_sentence:
         return {
@@ -76,8 +112,17 @@ async def translate_sentence_ai(sentence: str, target_word: str = "") -> Dict[st
             "cached": True,
         }
 
-    # 2. Local AI Generation (Ollama)
+    # 2. Retrieve offline dictionary hint to steer LLM towards authentic meanings
-    prompt = build_translation_prompt(clean_sentence, clean_word)
+    dict_hint = get_dictionary_hint(clean_word) if clean_word else ""
+
+    # 3. Local AI Generation (Ollama)
+    prompt = build_translation_prompt(
+        sentence=clean_sentence,
+        target_word=clean_word,
+        context=clean_context,
+        dict_hint=dict_hint,
+    )
+
     payload = {
         "model": config.OLLAMA_MODEL,
         "prompt": prompt,
@@ -85,7 +130,7 @@ async def translate_sentence_ai(sentence: str, target_word: str = "") -> Dict[st
         "format": "json",
         "options": {
             "temperature": 0.2,
-            "num_predict": 256,
+            "num_predict": 300,
         },
     }
 
@@ -94,17 +139,15 @@ async def translate_sentence_ai(sentence: str, target_word: str = "") -> Dict[st
     source = f"ollama:{config.OLLAMA_MODEL}"
 
     try:
-        async with httpx.AsyncClient(timeout=30.0) as client:
+        async with httpx.AsyncClient(timeout=35.0) as client:
             resp = await client.post(f"{config.OLLAMA_BASE_URL}/api/generate", json=payload)
             if resp.status_code == 200:
                 data = resp.json()
                 raw_response = data.get("response", "{}")
                 try:
                     parsed = json.loads(raw_response)
-                    sentence_ru = parsed.get("sentence_ru", "").strip()
-                    word_meaning = parsed.get("word_meaning", "").strip()
                 except Exception:
-                    # In case LLM emitted unescaped markdown or wrapper
+                    # In case LLM emitted markdown wrapper ```json ... ```
                     clean_raw = raw_response.strip()
                     if clean_raw.startswith("```json"):
                         clean_raw = clean_raw[7:]
@@ -112,15 +155,41 @@ async def translate_sentence_ai(sentence: str, target_word: str = "") -> Dict[st
                         clean_raw = clean_raw[3:]
                     if clean_raw.endswith("```"):
                         clean_raw = clean_raw[:-3]
+                    try:
                         parsed = json.loads(clean_raw.strip())
-                    sentence_ru = parsed.get("sentence_ru", "").strip()
+                    except Exception:
-                    word_meaning = parsed.get("word_meaning", "").strip()
+                        parsed = {}
+
+                sentence_ru = str(parsed.get("sentence_ru", "")).strip()
+
+                # Extract word meaning flexibly
+                raw_meaning = (
+                    parsed.get("word_meaning")
+                    or parsed.get(f"{clean_word}_meaning")
+                    or parsed.get("meaning")
+                    or ""
+                )
+                raw_meaning = str(raw_meaning).strip()
+
+                # Clean up format like "briefcase: портфель" or "портфель (сущ.)"
+                if ":" in raw_meaning:
+                    raw_meaning = raw_meaning.split(":", 1)[1].strip()
+
+                # If model returned the English term back, fallback to dictionary hint
+                if raw_meaning.lower() == clean_word.lower() or not raw_meaning:
+                    if dict_hint:
+                        raw_meaning = dict_hint.split(",")[0].strip()
+                    else:
+                        raw_meaning = ""
+
+                word_meaning = raw_meaning
+
             else:
                 logger.warning(f"Ollama returned HTTP {resp.status_code}: {resp.text}")
     except Exception as e:
         logger.warning(f"Ollama connection error: {e}")
 
-    # 3. Graceful Fallback if Ollama is unreachable
+    # 4. Graceful Fallback if Ollama is unreachable
     if not sentence_ru:
         sentence_ru = "(Локальный ИИ временно недоступен)"
         source = "fallback"

app/cefr.py

@@ -2056,10 +2056,13 @@ def resolve_context_sentence_for_term(
     if not clean_term:
         return ""
 
-    # 1. If candidate_sentence already contains this term (case-insensitive word boundary)
+    # 1. If candidate_sentence already contains this term (and is a complete sentence)
     if candidate_sentence:
         cand_clean = candidate_sentence.strip()
-        if re.search(rf"\b{re.escape(clean_term)}\b", cand_clean, re.IGNORECASE):
+        is_complete = bool(re.search(r"[.!?]['\"”’»)]*$", cand_clean)) and not bool(
+            re.search(r"\b(?:Mr|Mrs|Ms|Miss|Dr|Prof|St|Jr|Sr|[A-Za-z])\.['\"”’»)]*$", cand_clean, re.IGNORECASE)
+        )
+        if is_complete and re.search(rf"\b{re.escape(clean_term)}\b", cand_clean, re.IGNORECASE):
             return re.sub(r"\s+", " ", cand_clean.replace("___", clean_term)).strip()
 
     # 2. Look up in cefr_questions specifically for this term
@@ -2075,7 +2078,7 @@ def resolve_context_sentence_for_term(
 
     # 3. Look up in books table
     try:
-        books = conn.execute("SELECT chapters_json FROM books ORDER BY id DESC LIMIT 10").fetchall()
+        books = conn.execute("SELECT chapters_json FROM books ORDER BY id DESC LIMIT 50").fetchall()
         for b in books:
             try:
                 chapters = json.loads(b["chapters_json"])
@@ -2084,10 +2087,10 @@ def resolve_context_sentence_for_term(
             for ch in chapters:
                 content = ch.get("content", "")
                 if clean_term.lower() in content.lower():
-                    for s in re.split(r"(?<=[.!?])\s+", content):
+                    from app.miner import robust_sentence_split
-                        s_clean = s.strip()
+                    for s_clean in robust_sentence_split(content):
-                        if 20 <= len(s_clean) <= 180 and re.search(rf"\b{re.escape(clean_term)}\b", s_clean, re.IGNORECASE):
+                        if 20 <= len(s_clean) <= 240 and re.search(rf"\b{re.escape(clean_term)}\b", s_clean, re.IGNORECASE):
-                            return re.sub(r"\s+", " ", s_clean).strip()
+                            return s_clean.strip()
     except Exception:
         pass
 

app/config.py

@@ -42,5 +42,5 @@ DB_PATH = Path(os.getenv("DB_PATH", str(BASE_DIR / "data" / "notes.db")))
 DB_PATH.parent.mkdir(parents=True, exist_ok=True)
 
 OLLAMA_BASE_URL = os.getenv("OLLAMA_BASE_URL", "http://127.0.0.1:11434").rstrip("/")
-OLLAMA_MODEL = os.getenv("OLLAMA_MODEL", "qwen2.5:1.5b")
+OLLAMA_MODEL = os.getenv("OLLAMA_MODEL", "qwen2.5:3b")
 

app/database.py

@@ -386,7 +386,8 @@ def _migrate(conn: sqlite3.Connection):
 def repair_mismatched_cefr_contexts(conn: sqlite3.Connection):
     """
     Repairs legacy vocabulary cards that were imported with placeholder blanks ('___')
-    or with mismatched contexts from CEFR test questions.
+    or with truncated/mismatched contexts from CEFR test questions and reader.
+    Also purges truncated sentence cache entries.
     """
     try:
         from app.cefr import resolve_context_sentence_for_term
@@ -396,11 +397,25 @@ def repair_mismatched_cefr_contexts(conn: sqlite3.Connection):
         if not tbl:
             return
 
+        # Purge truncated sentence cache entries
+        trans_tbl = conn.execute("SELECT 1 FROM sqlite_master WHERE type='table' AND name='sentence_translations'").fetchone()
+        if trans_tbl:
+            conn.execute(
+                """
+                DELETE FROM sentence_translations
+                WHERE original_text LIKE '%pecked Mrs.%'
+                   OR original_text LIKE '%the John F.'
+                   OR instr(original_text, '___') > 0
+                """
+            )
+
         rows = conn.execute(
             """
             SELECT id, term, context_sentence
             FROM vocabulary_cards
             WHERE instr(context_sentence, '___') > 0
+               OR context_sentence LIKE '%pecked Mrs.%'
+               OR context_sentence LIKE '%the John F.'
             """
         ).fetchall()
 

app/main.py

@@ -190,6 +190,7 @@ class RfcFetchRequest(BaseModel):
 class SentenceTranslateRequest(BaseModel):
     sentence: str = Field(..., min_length=1)
     target_word: Optional[str] = ""
+    context: Optional[str] = ""
 
 
 class CefrAnswerRequest(BaseModel):
@@ -446,7 +447,7 @@ async def api_translate_sentence(
     clean_s = req.sentence.strip()
     if not clean_s:
         raise HTTPException(status_code=400, detail="Sentence cannot be empty")
-    return await translate_sentence_ai(clean_s, req.target_word or "")
+    return await translate_sentence_ai(clean_s, req.target_word or "", req.context or "")
 
 
 @app.get("/api/ai/status")

app/miner.py

@@ -102,31 +102,65 @@ for _lvl, _pos_map in CEFR_LEXICON.items():
                 _WORD_INDEX[_w_low] = []
             _WORD_INDEX[_w_low].append((_lvl, _pos))
 
-# Common abbreviations to prevent false splits
+def robust_sentence_split(text: str) -> List[str]:
-_HONORIFICS = ["Mr.", "Mrs.", "Ms.", "Dr.", "Prof.", "St.", "vs.", "etc.", "e.g.", "i.e.", "Jr.", "Sr."]
+    """
+    Splits English narrative prose into clean, complete sentences.
+    Protects decimals, acronyms, single-letter initials, honorifics, titles, and abbreviations.
+    """
+    if not text:
+        return []
+
+    # Step 1: Protect decimal numbers (e.g. 3.14, 1.5)
+    s = re.sub(r"(\d+)\.(\d+)", r"\1§DOT§\2", text)
+
+    # Step 2: Protect acronyms with dots like U.S., U.K., e.g., i.e., a.m., p.m., Ph.D.
+    s = re.sub(r"\b((?:[A-Za-z]\.){2,})", lambda m: m.group(1).replace(".", "§DOT§"), s)
+
+    # Step 3: Protect single-letter initials like 'John F. Kennedy', 'J. K. Rowling'
+    s = re.sub(r'\b([A-Za-z])\.(?=\s+[A-Za-z0-9"“\'‘«—])', r"\1§DOT§", s)
+
+    # Step 4: Protect common honorifics, titles, and abbreviations
+    abbr_pattern = (
+        r"\b("
+        r"Mr|Mrs|Ms|Miss|Dr|Prof|Rev|Hon|Capt|Col|Gen|Lt|Cmdr|Sgt|Cpl|Pvt|Maj|Adm|"
+        r"Gov|Sen|Rep|Pres|Sec|Treas|Insp|Det|Supt|Messrs|Mmes|St|Mt|Ft|"
+        r"etc|vs|v|Jr|Sr|approx|app|apt|dept|univ|assn|assoc|bldg|co|corp|inc|ltd|"
+        r"est|misc|al|ibid|op|cit|ed|eds|trans|vol|vols|no|nos|sec|secs|ch|chs|"
+        r"fig|figs|p|pp|ave|blvd|rd|st|ct|ln|dr|hwy|"
+        r"jan|feb|mar|apr|jun|jul|aug|sep|sept|oct|nov|dec"
+        r")\."
+    )
+    s = re.sub(abbr_pattern, r"\1§DOT§", s, flags=re.IGNORECASE)
+
+    # Step 5: Protect ellipsis (... or ....)
+    s = re.sub(r"\.{2,}", lambda m: "§ELLIPSIS§" + ("§DOT§" * len(m.group(0))), s)
+
+    # Step 6: Protect any period followed by lowercase letter
+    s = re.sub(r"\.([\'\"”’»)]*\s+[a-zа-яё])", r"§DOT§\1", s)
+
+    # Step 7: Split on sentence terminators [.!?] followed by whitespace and uppercase/number/quote
+    marked = re.sub(r'([.!?][\'\"”’»)]*)(\s+)(?=[A-Z0-9"“\'‘«—]|$)', r"\1§SEP§", s)
+
+    raw_sentences = [
+        re.sub(r"\s+", " ", part.replace("§DOT§", ".").replace("§ELLIPSIS§", "")).strip()
+        for part in marked.split("§SEP§")
+    ]
+    return [sent for sent in raw_sentences if sent]
 
 
 def split_into_clean_sentences(text: str) -> List[str]:
     """
     Safely splits narrative prose into clean, grammatically complete sentences.
-    Protects honorifics and abbreviations from being split incorrectly.
+    Protects honorifics, initials, and abbreviations from being split incorrectly.
     """
     if not text:
         return []
 
-    # Protect common abbreviations and titles
+    raw_splits = robust_sentence_split(text)
-    protected = text
-    for abbr in _HONORIFICS:
-        protected = protected.replace(abbr, abbr.replace(".", "§DOT§"))
-
-    # Split on sentence-ending punctuation (.!?) followed by space or newline
-    raw_splits = re.split(r"(?<=[.!?])\s+", protected)
 
     clean_sentences = []
     for s in raw_splits:
-        restored = s.replace("§DOT§", ".").strip()
+        flattened = re.sub(r"\s+", " ", s).strip()
-        # Flatten multiple spaces and line breaks
-        flattened = re.sub(r"\s+", " ", restored)
 
         # Basic validity filters:
         # 1. Skip dialogue lines starting with quotation marks or dashes

app/static/js/modules/api.js

@@ -114,13 +114,14 @@ export const api = {
             const resp = await fetch(`/api/detect-phrasal-verb?sentence=${encodeURIComponent(sentence)}&word=${encodeURIComponent(word)}`);
             return resp.ok ? await resp.json() : { detected: false, phrasal_verb: null };
         },
-        async translateSentence(sentence, targetWord = '') {
+        async translateSentence(sentence, targetWord = '', context = '') {
             const resp = await fetch('/api/translate-sentence', {
                 method: 'POST',
                 headers: { 'Content-Type': 'application/json' },
                 body: JSON.stringify({
                     sentence: sentence,
-                    target_word: targetWord
+                    target_word: targetWord,
+                    context: context
                 })
             });
             if (resp.ok) {

app/static/js/modules/reader.js

@@ -1104,40 +1104,51 @@ export function renderInteractiveSpans(container, text, contextText, knownWordsM
     });
 }
 
-const READER_HONORIFICS = [
+export function robustSentenceSplit(text) {
-    'Mr.', 'Mrs.', 'Ms.', 'Dr.', 'Prof.', 'St.', 'vs.', 'etc.', 'e.g.', 'i.e.', 'Jr.', 'Sr.',
+    if (!text) return [];
-    'Capt.', 'Col.', 'Gen.', 'Lt.', 'Sgt.', 'Rev.', 'Hon.'
+    let s = text;
-];
+
+    // Step 1: Protect decimal numbers (e.g. 3.14, 1.5)
+    s = s.replace(/(\d+)\.(\d+)/g, '$1§DOT§$2');
+
+    // Step 2: Protect acronyms with dots like U.S., U.K., e.g., i.e., a.m., p.m., Ph.D.
+    s = s.replace(/\b((?:[A-Za-z]\.){2,})/g, (m) => m.replace(/\./g, '§DOT§'));
+
+    // Step 3: Protect single-letter initials like 'John F. Kennedy', 'J. K. Rowling'
+    s = s.replace(/\b([A-Za-z])\.(?=\s+[A-Za-z0-9"“'‘«—])/g, '$1§DOT§');
+
+    // Step 4: Protect common honorifics, titles, and abbreviations
+    const abbrPattern = /\b(Mr|Mrs|Ms|Miss|Dr|Prof|Rev|Hon|Capt|Col|Gen|Lt|Cmdr|Sgt|Cpl|Pvt|Maj|Adm|Gov|Sen|Rep|Pres|Sec|Treas|Insp|Det|Supt|Messrs|Mmes|St|Mt|Ft|etc|vs|v|Jr|Sr|approx|app|apt|dept|univ|assn|assoc|bldg|co|corp|inc|ltd|est|misc|al|ibid|op|cit|ed|eds|trans|vol|vols|no|nos|sec|secs|ch|chs|fig|figs|p|pp|ave|blvd|rd|st|ct|ln|dr|hwy|jan|feb|mar|apr|jun|jul|aug|sep|sept|oct|nov|dec)\./gi;
+    s = s.replace(abbrPattern, '$1§DOT§');
+
+    // Step 5: Protect ellipsis (... or ....)
+    s = s.replace(/\.{2,}/g, (m) => '§ELLIPSIS§' + '§DOT§'.repeat(m.length));
+
+    // Step 6: Protect any period followed by lowercase letter (e.g. 'word. otherword')
+    s = s.replace(/\.([\'"”’»)]*\s+[a-zа-яё])/g, '§DOT§$1');
+
+    // Step 7: Split on sentence terminators [.!?] followed by whitespace and uppercase/number/quote/start of dialog
+    const marked = s.replace(/([.!?]['"”’»)]*)(\s+)(?=[A-Z0-9"“'‘«—]|$)/g, '$1§SEP§');
+
+    return marked.split('§SEP§')
+        .map(part => part.replace(/§DOT§/g, '.').replace(/§ELLIPSIS§/g, '').trim())
+        .filter(Boolean);
+}
 
 export function extractContextSentence(fullText, word) {
     if (!fullText) return '';
-    let protectedText = fullText;
+    const sentences = robustSentenceSplit(fullText);
-    for (const h of READER_HONORIFICS) {
-        if (protectedText.toLowerCase().includes(h.toLowerCase())) {
-            const safe = h.replace(/\./g, '§DOT§');
-            const reH = new RegExp(escapeRegex(h), 'gi');
-            protectedText = protectedText.replace(reH, safe);
-        }
-    }
-    // Protect decimal numbers (e.g. 3.14, 1.5)
-    protectedText = protectedText.replace(/(\d+)\.(\d+)/g, '$1§DOT§$2');
-
-    // Split sentences on terminal punctuation (.!?) followed by optional quotes/brackets and whitespace
-    const marked = protectedText.replace(/([.!?]['"”’»)]*)\s+/g, '$1§SEP§');
-    const rawSentences = marked.split('§SEP§');
-
     const cleanWord = (word || '').trim();
     if (cleanWord) {
         const escaped = escapeRegex(cleanWord);
         const wordRe = new RegExp(`(?:^|\\W)${escaped}(?:$|\\W)`, 'i');
-        for (const s of rawSentences) {
+        for (const s of sentences) {
             if (wordRe.test(s)) {
-                return s.replace(/§DOT§/g, '.').trim();
+                return s.replace(/\s+/g, ' ').trim();
             }
         }
     }
-
+    return sentences.length > 0 ? sentences[0].replace(/\s+/g, ' ').trim() : fullText.slice(0, 160).trim();
-    return fullText.slice(0, 160).trim();
 }
 
 // -----------------------------------------------------------------------
@@ -1176,7 +1187,7 @@ export function positionWordPopup(targetElement) {
     readerWordPopup.style.left = `${Math.round(left)}px`;
 }
 
-export async function openWordPopup(term, contextSentence, targetElement) {
+export async function openWordPopup(term, contextSentence, targetElement, paragraphContext = '') {
     const readerWordPopup = document.getElementById('reader-word-popup');
     const readerSettingsPanel = document.getElementById('reader-settings-panel');
     const popupTerm = document.getElementById('popup-term');
@@ -1202,6 +1213,7 @@ export async function openWordPopup(term, contextSentence, targetElement) {
     popupCurrentData = {
         term: cleanTerm,
         contextSentence: contextSentence,
+        paragraphContext: paragraphContext,
         targetElement: targetElement,
         existingCard: existingCard,
     };
@@ -1339,7 +1351,7 @@ export async function openWordPopup(term, contextSentence, targetElement) {
 
 export async function handleWordClick(word, paragraphText, targetElement) {
     const sentence = extractContextSentence(paragraphText, word);
-    openWordPopup(word, sentence, targetElement);
+    openWordPopup(word, sentence, targetElement, paragraphText);
 }
 
 function updatePopupPositionOnScroll() {
@@ -1965,7 +1977,7 @@ export function initReader() {
             if (popupContextMeaningWrap) popupContextMeaningWrap.classList.add('hidden');
 
             try {
-                const res = await api.lookup.translateSentence(ctxSentence, term);
+                const res = await api.lookup.translateSentence(ctxSentence, term, popupCurrentData.paragraphContext || '');
                 if (popupSentenceLoading) popupSentenceLoading.classList.add('hidden');
                 if (popupSentenceTranslationWrap) popupSentenceTranslationWrap.classList.remove('hidden');
 

app/static/sw.js

@@ -1,7 +1,7 @@
 // Service Worker for Notes English PWA (Stage 2)
 // Strategy: Cache-first for static assets, Network-first for API calls.
 
-const CACHE_NAME = 'notes-v5.9';
+const CACHE_NAME = 'notes-v6.0';
 const OFFLINE_URL = '/';
 
 const PRECACHE_URLS = [

docker-compose.yml

@@ -16,6 +16,7 @@ services:
       - ADMIN_USERNAME=kopri
       - DB_PATH=/app/data/notes.db
       - OLLAMA_BASE_URL=http://ollama:11434
+      - OLLAMA_MODEL=qwen2.5:3b
     depends_on:
       - ollama
 

tests/test_api.py

@@ -1867,6 +1867,94 @@ def test_cefr_context_resolution_and_repair():
         assert "negotiate" in repaired["context_sentence"].lower()
 
 
+def test_robust_sentence_split_and_ai_translation(monkeypatch):
+    from app.miner import robust_sentence_split
+    from app.ai import build_translation_prompt
+
+    # 1. Test robust sentence segmentation
+    # Case A: Initials in names (John F. Kennedy)
+    s_jfk = "Perhaps the most famous is the John F. Kennedy International Airport in New York, but there is also the John F. Kennedy Space Center."
+    splits_jfk = robust_sentence_split(s_jfk)
+    assert len(splits_jfk) == 1
+    assert "John F. Kennedy International Airport" in splits_jfk[0]
+
+    # Case B: Honorifics with dots (Mr., Mrs., Dr., etc.)
+    s_dursley = "When Mr. and Mrs. Dursley woke up on the dull, gray Tuesday our story starts, there was nothing about the cloudy sky outside to suggest that strange things would happen."
+    splits_dursley = robust_sentence_split(s_dursley)
+    assert len(splits_dursley) == 1
+    assert "Mr. and Mrs. Dursley" in splits_dursley[0]
+
+    # Case C: Abbreviations (No. 4, U.S. Navy)
+    s_no4 = "They lived at No. 4 Privet Drive, which was a very quiet street. Next sentence starts here."
+    splits_no4 = robust_sentence_split(s_no4)
+    assert len(splits_no4) == 2
+    assert splits_no4[0] == "They lived at No. 4 Privet Drive, which was a very quiet street."
+    assert splits_no4[1] == "Next sentence starts here."
+
+    s_us = "The men thought that the U.S. Navy would send a plane to look for them. They waited patiently."
+    splits_us = robust_sentence_split(s_us)
+    assert len(splits_us) == 2
+    assert "U.S. Navy" in splits_us[0]
+
+    # Case D: Complex paragraph with Mr., Mrs., dialogue
+    s_hp = (
+        "At half past eight, Mr. Dursley picked up his briefcase, pecked Mrs. Dursley on the cheek, "
+        "and tried to kiss Dudley good-bye but missed, because Dudley was now having a tantrum and "
+        "throwing his cereal at the walls. “Little tyke,” chortled Mr. Dursley as he left the house."
+    )
+    splits_hp = robust_sentence_split(s_hp)
+    assert len(splits_hp) == 2
+    assert "pecked Mrs. Dursley on the cheek" in splits_hp[0]
+    assert splits_hp[0].startswith("At half past eight, Mr. Dursley")
+    assert "Little tyke" in splits_hp[1]
+
+    # 2. Test build_translation_prompt
+    prompt = build_translation_prompt(
+        sentence="At half past eight, Mr. Dursley picked up his briefcase.",
+        target_word="briefcase",
+        context="Full context paragraph here.",
+        dict_hint="портфель, кейс, дипломат",
+    )
+    assert "Контекст абзаца:" in prompt
+    assert "Full context paragraph here." in prompt
+    assert "briefcase" in prompt
+    assert "портфель, кейс, дипломат" in prompt
+    assert "Mr. -> мистер" in prompt
+    assert "НЕ дословным, без топорного калькирования" in prompt
+
+    # 3. Test API endpoint /api/translate-sentence
+    # Authenticate client
+    login_resp = client.post("/api/login", json={"username": "testadmin", "password": "testpass12345"})
+    assert login_resp.status_code == 200
+
+    # Mock translate_sentence_ai
+    import app.main as main_mod
+
+    async def mock_translate_sentence_ai(sentence, target_word="", context=""):
+        return {
+            "original_sentence": sentence,
+            "sentence_ru": "В половине девятого мистер Дурсли взял свой портфель.",
+            "target_word": target_word,
+            "word_meaning": "портфель",
+            "source": "mock_ollama",
+            "cached": False,
+        }
+
+    monkeypatch.setattr(main_mod, "translate_sentence_ai", mock_translate_sentence_ai)
+
+    tr_resp = client.post("/api/translate-sentence", json={
+        "sentence": "At half past eight, Mr. Dursley picked up his briefcase.",
+        "target_word": "briefcase",
+        "context": "Context paragraph",
+    })
+    assert tr_resp.status_code == 200
+    tr_data = tr_resp.json()
+    assert tr_data["sentence_ru"] == "В половине девятого мистер Дурсли взял свой портфель."
+    assert tr_data["word_meaning"] == "портфель"
+    assert tr_data["target_word"] == "briefcase"
+
+
+