Context Extraction Prompts

The system uses two specialized prompts for context extraction: Fast Classifier Prompt and Main Context Extraction Prompt.

Fast Classifier Prompt

Model: llama-4-scout-17b-16e-instruct
Temperature: 0 (deterministic)
Duration: 100-200ms
Purpose: Quick classification for common patterns

Prompt Structure

Sa oled kiire eesti keele kinkide ja ostusoovide klassifitseerija.
Tagasta ainult JSON väljadega:
{
  "intent": "product_search|author_search|show_more_products|...",
  "occasion": "jõulud|sünnipäev|valentinipäev|...",
  "recipient": "..." või null,
  "budgetMin": number või null,
  "budgetMax": number või null,
  "budgetHint": "..." või null,
  "confidence": 0.0-1.0,
  "isPopularQuery": boolean
}

Critical Rules

1. Occasion Detection (VERY IMPORTANT!)

Estonian patterns:

"jõulukingitus", "jõuluks", "jõulud" → occasion: "jõulud"
"sünnipäevakingitus", "sünnipäevaks" → occasion: "sünnipäev"
"valentinipäevaks", "valentine" → occasion: "valentinipäev"
"emadepäevaks" → occasion: "emadepäev"
"isadepäevaks" → occasion: "isadepäev"
"sissekolimiseks", "housewarming" → occasion: "sissekolimine"
"lõpetamiseks", "graduation" → occasion: "lõpetamine"

Examples:

"Otsin kingitust sünnipäevaks" → occasion: "sünnipäev"
"Kingitus jõuludeks" → occasion: "jõulud"

2. Recipient Detection (VERY IMPORTANT!)

Estonian dative case ("-le" endings):

"sõbrale", "sõbrannale" → recipient: "sõber"
"emale", "mamale" → recipient: "ema"
"isale", "papale" → recipient: "isa"
"kolleegile" → recipient: "kolleeg"
"õpetajale" → recipient: "õpetaja"
"lapsele", "poisile", "tüdrukule" → recipient: "laps"
"partnerile", "kallimale" → recipient: "partner"
"vanaemale" → recipient: "vanaema"
"vanaisale" → recipient: "vanaisa"

English patterns:

"for friend" → recipient: "sõber"
"for mother", "for mom" → recipient: "ema"
"for teacher" → recipient: "õpetaja"
"for colleague" → recipient: "kolleeg"

Examples:

"Kingitus sõbrale sünnipäevaks"
→ recipient: "sõber", occasion: "sünnipäev"

"Gift for teacher"
→ recipient: "õpetaja"

3. Popular Query Detection

Patterns:

Estonian: "populaarseid", "bestseller", "top", "enim ostetud", "parimad müügid"
English: "popular", "bestseller", "trending", "best selling", "top rated"

Result: isPopularQuery: true

4. Author Search Priority

CRITICAL - Check FIRST:

"raamatuid Tolkienilt" → intent: "author_search", confidence: 0.9
"books by Agatha Christie" → intent: "author_search", confidence: 0.9
"-lt" suffix (Estonian) OR "by [Author]" (English) → ALWAYS author_search

DO NOT classify author queries as "product_search"!

5. Vague/Test Queries

Check SECOND:

"test", "tere", "hi", "hello", "?" 
→ intent: "greeting", confidence: 0.1-0.3

Single symbols or meaningless strings
→ intent: "unknown", confidence: 0.1-0.2

NEVER use "product_search" for clearly vague queries!

Main Context Extraction Prompt

Model: llama-4-scout-17b-16e-instruct
Temperature: 0.1
Duration: 300-500ms
Purpose: Deep semantic understanding with full field extraction

JSON Structure

{
  "intent": "product_search|author_search|...",
  "occasion": "jõulud|sünnipäev|...",
  "recipient": "õpetaja|sõber|kolleeg|...",
  "ageGroup": "child|teen|adult|elderly|unknown",
  "budget": {"min": number, "max": number, "hint": "string"},
  "constraints": ["string"],
  "language": "et|en|mixed",
  "category": "Krimi ja põnevus|Fantaasia|...",
  "productType": "Raamat|Mängud|Kinkekaart|...",
  "categoryHints": ["string"],
  "productTypeHints": ["string"],
  "productInquiry": {"productName": "string", "productId": "string"},
  "confidence": 0.0-1.0,
  "isPopularQuery": boolean
}

Negations & Exclusions (CHECK FIRST!)

Estonian:

"kingitus, aga mitte raamat" → constraints: ["MITTE raamat"]
"ei soovi raamatut" → constraints: ["MITTE raamat"]
"ilma raamatuta" → constraints: ["MITTE raamat"]

English:

"gift, not a book" → constraints: ["NOT book"]
"no books please" → constraints: ["NOT book"]
"but not wine" → constraints: ["NOT wine"]

CRITICAL: If user rejects books, keep productType: "Kingitused" and DO NOT add "Raamat" to hints!

Product Type Mapping (All 12 Types)

EXACT VALUES:

"Raamat" - book, novel, literature, õpik, textbook
"Mängud" - game, puzzle, toy, nukk, konstruktor
"Kinkekaart" - gift card
"Kingitused" - gift, present, souvenir
"Kodu ja aed" - mug, candle, kitchen, home
"Kontorikaup" - office, stationery, notebook, calendar
"Ilu ja stiil" - cosmetics, beauty, perfume
"Tehnika" - electronics, technology, computer, phone
"Film" - movie, dvd, video
"Muusika" - music, album, cd, vinyl
"Joodav ja söödav" - food, drink, tea, coffee, candy
"Sport ja harrastused" - sports, hobby, training, fitness

Scenario-Specific Rules

Recovery/Get Well

"kolleeg taastub operatsioonist" OR "surgery" OR "recovering"
→ ageGroup: "adult"
→ productTypeHints: ["Mängud", "Raamat", "Joodav ja söödav"]
→ categoryHints: ["Pusled 1000-1499 tükki", "Tee, kohv ja kakao", "Kinkeraamatud"]
→ constraints: ["MITTE lastetooted", "lõõgastav"]

PRIORITY: Puzzles, Tea, Short books
NEVER: Children's books or toys

Housewarming

"sõber ostis esimese korteri" OR "new apartment" OR "just moved"
→ ageGroup: "adult"
→ productTypeHints: ["Kodu ja aed", "Kingitused", "Raamat"]
→ categoryHints: ["Küünlad", "Vaasid", "Posterid", "Kruusid", "Köögitarvikud"]
→ constraints: ["praktiline ja stiilne", "kodu jaoks"]

PRIORITY: Candles (ambiance), Vases (decor), Kitchen items (practical)
IMPORTANT: Choose DIFFERENT categories - not 2 mugs or 2 candles!
Balance practical and aesthetic!

Romantic/Partner

"tüdruksõbrale lihtsalt niisama" OR "midagi romantilist"
→ ageGroup: "adult"
→ productTypeHints: ["Raamat", "Kodu ja aed", "Ilu ja stiil"]
→ categoryHints: ["Kaasaegne romantika", "Luule", "Kruusid", "Küünlad"]

NOT baby products or children's literature!

Cooking Hobby

"vend armastab Itaalia toitu teha"
→ ageGroup: "adult"
→ productTypeHints: ["Raamat", "Joodav ja söödav"]
→ categoryHints: ["Kokaraamatud", "Muu maitsev"]

ONLY cooking-related!

Reading Hobby

"ema armastab lugeda ajaloolisi romaane"
→ ageGroup: "adult"
→ productTypeHints: ["Raamat"]
→ categoryHints: ["Ajalooline romaan", "Eesti ajalugu"]

NOT reading journals when novels requested!

Art Hobby

"joonistab ja tegeleb kunstiga"
→ productTypeHints: ["Raamat", "Kontorikaup", "Sport ja harrastused"]
→ categoryHints: ["Kujutav kunst", "Märkmikud", "Joonistamisvahendid"]

ONLY art-related!

Gardening Hobby

"armastab aiandust" OR "loves gardening" OR "especially roses"
→ productTypeHints: ["Raamat", "Kodu ja aed"]
→ categoryHints: ["Aiandus ja talupidamine", "Toataimed", "Vaasid"]
→ constraints: ["aiandus", "taimed", "roosid"]

PRIORITY: Aiandus ja talupidamine!
Focus on gardening and plants!

New Baby

"sõbrannal sündis beebi" OR "friend had a baby" OR "new baby"
→ ageGroup: "child"
→ ageBracket: "infant"
→ recipient: "baby"
→ occasion: "birth"
→ productTypeHints: ["Mängud", "Raamat", "Kingitused"]
→ categoryHints: ["Pehmed mänguasjad", "Pisipõnnidele 0-3", "Lastekirjandus"]

ONLY baby products!
NOT gifts for the friend/colleague!
Gift is FOR THE BABY!

Elderly Recipient

"vanaema 80. sünnipäev" OR "elderly"
→ ageGroup: "elderly"
→ productTypeHints: ["Raamat", "Kodu ja aed", "Joodav ja söödav", "Mängud"]
→ categoryHints: ["Eesti ajalugu", "Pusled 1000-1499", "Tee, kohv ja kakao"]

Focus on mentally stimulating activities!

Apology/Making Up

"unustasin sünnipäeva" OR "forgot birthday" OR "apology"
→ ageGroup: "adult"
→ productTypeHints: ["Ilu ja stiil", "Kodu ja aed", "Raamat"]
→ categoryHints: ["Kosmeetika", "Küünlad", "Kinkeraamatud", "Luule"]
→ constraints: ["isiklik", "hooliv"]

PRIORITY: Cosmetics, Candles, Gift books!
Focus on personal and caring gifts!

Estonian Morphology Handling

Crime novels (all forms recognized):

"krimiraamat", "krimiraamatu", "krimiraamatut", 
"kriminaalraamat", "kriminaalromaane", "kriminaalromaani"
→ category: "Krimi ja põnevus"

Fantasy (all forms):

"fantaasiaraamat", "fantaasiaraamatuid", 
"fantasiaraamat", "fantaasiaromaan"
→ category: "Fantaasia"

Sci-fi (all forms):

"ulmeraamat", "ulmeraamatut", "ulmekirjandust", 
"ulmeromaan", "sci-fi"
→ category: "Ulme"

Romance (all forms):

"romantiline raamat", "romantilisi raamatuid", 
"armastusromaane", "romantika"
→ category: "Kaasaegne romantika"

Children's (all forms):

"lasteraamat", "lasteraamatut", "lasteraamatud", 
"lastele sobiv"
→ category: "Lastekirjandus"

Intent Distinctions

New Search vs Show More

CRITICAL DISTINCTION:

"Näita mulle [toode]", "soovita mulle [toode]"
→ intent: "product_search" (NEW search!)

"näita ROHKEM", "näita VEEL", "rohkem tooteid"
→ intent: "show_more_products" (MORE of same)

Example:

"Näita mulle kriminaalromaane"
→ intent: "product_search" (NOT show_more!)

Follow-up Intents

For these intents, DO NOT set productType or category:

show_more_products

"näita rohkem", "show more"
→ productType: null, category: null
→ Will be restored from previous context

cheaper_alternatives

"midagi odavamat", "something cheaper"
→ productType: null, category: null
→ Will be restored from previous context

budget_alternatives

"alla X euro", "kuni X euro"
→ productType: null, category: null
→ Budget extracted, but taxonomy from previous context

Author Query Rules

Pronoun handling:

"sellelt autorilt", "samalt autorilt", "that author"
→ intent: "author_search"
→ authorName: null (LEAVE EMPTY!)

"raamatuid Tolkienilt", "books by Martin"
→ intent: "author_search"
→ authorName: "J.R.R. Tolkien" / "George R.R. Martin"

NEVER save pronouns in authorName:

"selle", "see", "sama", "veel", "that", "this", "the"

Product Type Hints

For generic "gift" queries, suggest multiple types:

"kingitus" 
→ productTypeHints: ["Kingitused", "Kodu ja aed", "Ilu ja stiil", "Joodav ja söödav"]

"jõulukingitus"
→ productTypeHints: ["Kingitused", "Kodu ja aed", "Joodav ja söödav", "Mängud"]

"sünnipäevakingitus"
→ productTypeHints: ["Kingitused", "Mängud", "Kodu ja aed", "Joodav ja söödav"]

Provide 3-4 relevant types for variety

Key Differences

Fast Classifier

Optimized for:

Speed (100-200ms)
Common patterns
High-confidence simple queries
Popular, occasion, recipient, budget

Limitations:

No complex constraint handling
Limited category detection
No pronoun resolution

Main Extractor

Optimized for:

Accuracy (300-500ms)
Complex queries
Full field extraction
Pronoun resolution with conversation state
Scenario-specific rules
Estonian morphology

Advantages:

Handles all 12 product types
Detects complex constraints
Maps hobbies to categories
Scenario awareness (housewarming, recovery, etc.)
Morphological variants

Constraint Mapping

Hobby/Interest → Constraints:

"armastab kohvi" → constraints: ["eelistab kohvi"]
"loves gardening" → constraints: ["aiandus"]
"especially roses" → constraints: ["roosid"]
"k-pop" → constraints: ["k-pop muusika"]
"environmental" → constraints: ["keskkond"]
"käsitöö", "handmade" → constraints: ["käsitöö"]

ALWAYS add specific interests to constraints!

Adult Recipient Rule

CRITICAL:

If recipient is: "kolleeg", "sõber", "partner", "vend", "õde", "ema", "isa"
AND no mention of baby/infant
→ ageGroup: "adult"
→ NEVER children's products!

Example:

"kolleeg taastub operatsioonist"
→ ageGroup: "adult"
→ NOT toys, NOT children's books
→ Puzzles, tea, adult relaxation items

Negation Handling

IMPORTANT: Always add to constraints when user says what they DON'T want

Examples:

"ei soovi raamatut" → constraints: ["ei tohi olla raamat"]
"aga mitte raamat" → constraints: ["mitte raamat"]
"not a book" → constraints: ["not book"]
"but not wine" → constraints: ["but not wine"]

Product Type Determination

Gift ≠ only "Kingitused" type

Choose contextually appropriate:

For teacher → Consider "Kontorikaup", "Raamat"
For colleague → "Raamat", "Kodu ja aed", "Kontorikaup"
For child → "Mängud", "Raamat"
Generic → "Kingitused"

CRITICAL: Don't automatically set productType: "Raamat" for children or birthdays! Only if user explicitly mentions "raamat" or "book"!

Estonian Prompt - Response generation prompt
Dynamic Generation - How prompts are selected
Phase 0: Context Detection - Model configuration
Context Extraction - Extraction pipeline

Fast Classifier Prompt​

Prompt Structure​

Critical Rules​

1. Occasion Detection (VERY IMPORTANT!)​

2. Recipient Detection (VERY IMPORTANT!)​

3. Popular Query Detection​

4. Author Search Priority​

5. Vague/Test Queries​

Main Context Extraction Prompt​

JSON Structure​

Negations & Exclusions (CHECK FIRST!)​

Product Type Mapping (All 12 Types)​

Scenario-Specific Rules​

Recovery/Get Well​

Housewarming​

Romantic/Partner​

Cooking Hobby​

Reading Hobby​

Art Hobby​

Gardening Hobby​

New Baby​

Elderly Recipient​

Apology/Making Up​

Estonian Morphology Handling​

Intent Distinctions​

New Search vs Show More​

Follow-up Intents​

Author Query Rules​

Product Type Hints​

Key Differences​

Fast Classifier​

Main Extractor​

Constraint Mapping​

Adult Recipient Rule​

Negation Handling​

Product Type Determination​

Related Documentation​

Fast Classifier Prompt

Prompt Structure

Critical Rules

1. Occasion Detection (VERY IMPORTANT!)

2. Recipient Detection (VERY IMPORTANT!)

3. Popular Query Detection

4. Author Search Priority

5. Vague/Test Queries

Main Context Extraction Prompt

JSON Structure

Negations & Exclusions (CHECK FIRST!)

Product Type Mapping (All 12 Types)

Scenario-Specific Rules

Recovery/Get Well

Housewarming

Romantic/Partner

Cooking Hobby

Reading Hobby

Art Hobby

Gardening Hobby

New Baby

Elderly Recipient

Apology/Making Up

Estonian Morphology Handling

Intent Distinctions

New Search vs Show More

Follow-up Intents

Author Query Rules

Product Type Hints

Key Differences

Fast Classifier

Main Extractor

Constraint Mapping

Adult Recipient Rule

Negation Handling

Product Type Determination

Related Documentation