Nov 4, 2025 | LLM (AI), Pozivi (prompti)
Dobro izdelani pozivi (prompts) zmanjšajo halucinacije v velikih jezikovnih modelih (LLM), kjer model generira zavajajoče ali nepreverljive informacije.
Predstavljenih je sedem ključnih tehnik za inženiring pozivov s praktičnimi predlogami, uporabnimi za samostojne modele LLM in sisteme z razširjenim pridobivanjem podatkov (RAG).
-
Spodbujanje zadržanosti in odgovorov “Ne vem”:
Pozivi naj modelu dovolijo, da reče “Ne vem”, če ni prepričan—tako se izognemo izmišljenim odgovorom.
“You are a fact-checking assistant. If you are not confident in an answer, respond: ‘I don’t have enough information to answer that.’ If confident, give your answer with a short justification.”
-
Strukturirano, korak-po-korak razmišljanje:
Zahtevajte od modela, naj odgovarja z algoritmičnim, postopnim razmišljanjem za večjo doslednost.
“Please think through this problem step by step:
1) What information is given?
2) What assumptions are needed?
3) What conclusion follows logically?”
-
Utemeljevanje z “Po podatkih…”:
Pozivi naj zahtevajo sklicevanje na imenovane, verodostojne vire—manj izmišljotin, več dejstev.
“According to the World Health Organization (WHO) report from 2023, explain the main drivers of antimicrobial resistance. If the report doesn’t provide enough detail, say ‘I don’t know.’”
-
RAG z eksplicitnimi navodili in kontekstom:
Pozivi naj model striktno omejijo na podatke, ki jih pridobi iz virov, in naj izpostavijo manjkajoče informacije.
*[Assume two retrieved documents: X and Y]*
“Using only the information in X and Y, summarize the main causes of deforestation in the Amazon basin and related infrastructure projects. If the documents don’t cover a point, say ‘insufficient data.’”
-
Omejitve izhoda in omejevanje obsega:
Določite format, dolžino ali vsebinski obseg izhoda, da preprečite nepotrebne podrobnosti ali špekulacije.
“In no more than 100 words, summarize the role of mitochondria in human cells. If unsure, respond ‘I don’t know.’”
-
RAG-verifikacijska zanka (Chain-of-Verification):
Model naj najprej generira odgovor, nato preveri podatke v virih in po potrebi popravi odgovor z navedbo vira.
“Step 1: Generate an initial answer to the question: ‘When did the Berlin Wall fall?’
Step 2: Retrieve and read relevant passages from a trusted historical database.
Step 3: Compare the retrieved evidence to your answer.
Step 4: If discrepancies exist, correct the answer and cite the retrieved source.”
-
Specifični pozivi, opozorila in varnostne omejitve:
Za področja z visokim tveganjem (npr. medicina) uporabite strogo omejene pozive, zahtevajte citate in vklopite opozorila.
“You are a certified medical information assistant. Using peer-reviewed studies or official guidelines published before 2024, explain the first-line treatment for moderate persistent asthma in adults. If you cannot cite such a guideline, respond: ‘I cannot provide a recommendation; consult a medical professional.’”
Oct 15, 2025 | Pozivi (prompti)
https://cookbook.openai.com/examples/sora/sora2_prompting_guide
Primer referenčnega prompta za generiranje videa:
Format & Look
Duration 4s; 180° shutter; digital capture emulating 65 mm photochemical contrast; fine grain; subtle halation on speculars; no gate weave.
Lenses & Filtration
32 mm / 50 mm spherical primes; Black Pro-Mist 1/4; slight CPL rotation to manage glass reflections on train windows.
Grade / Palette
Highlights: clean morning sunlight with amber lift.
Mids: balanced neutrals with slight teal cast in shadows.
Blacks: soft, neutral with mild lift for haze retention.
Lighting & Atmosphere
Natural sunlight from camera left, low angle (07:30 AM).
Bounce: 4×4 ultrabounce silver from trackside.
Negative fill from opposite wall.
Practical: sodium platform lights on dim fade.
Atmos: gentle mist; train exhaust drift through light beam.
Location & Framing
Urban commuter platform, dawn.
Foreground: yellow safety line, coffee cup on bench.
Midground: waiting passengers silhouetted in haze.
Background: arriving train braking to a stop.
Avoid signage or corporate branding.
Wardrobe / Props / Extras
Main subject: mid-30s traveler, navy coat, backpack slung on one shoulder, holding phone loosely at side.
Extras: commuters in muted tones; one cyclist pushing bike.
Props: paper coffee cup, rolling luggage, LED departure board (generic destinations).
Sound
Diegetic only: faint rail screech, train brakes hiss, distant announcement muffled (-20 LUFS), low ambient hum.
Footsteps and paper rustle; no score or added foley.
Optimized Shot List (2 shots / 4 s total)
0.00–2.40 — “Arrival Drift” (32 mm, shoulder-mounted slow dolly left)
Camera slides past platform signage edge; shallow focus reveals traveler mid-frame looking down tracks. Morning light blooms across lens; train headlights flare softly through mist. Purpose: establish setting and tone, hint anticipation.
2.40–4.00 — “Turn and Pause” (50 mm, slow arc in)
Cut to tighter over-shoulder arc as train halts; traveler turns slightly toward camera, catching sunlight rim across cheek and phone screen reflection. Eyes flick up toward something unseen. Purpose: create human focal moment with minimal motion.
Camera Notes (Why It Reads)
Keep eyeline low and close to lens axis for intimacy.
Allow micro flares from train glass as aesthetic texture.
Preserve subtle handheld imperfection for realism.
Do not break silhouette clarity with overexposed flare; retain skin highlight roll-off.
Finishing
Fine-grain overlay with mild chroma noise for realism; restrained halation on practicals; warm-cool LUT for morning split tone.
Mix: prioritize train and ambient detail over footstep transients.
Poster frame: traveler mid-turn, golden rim light, arriving train soft-focus in background haze.
May 12, 2025 | LLM (AI), Pozivi (prompti)
ChatGPT & Grok sta dobila dostop do praktično “neomejenega pomnilnika”.
☑ Zapomnita si vaše pretekle pogovore.
☑ Oba lahko naučite, kdo ste in kaj želite.
Kako dodate trajne podatke v “svoj” spomin:
1 – Pojdite na ChatGPT ali Grok. Ne deluje na Gemini & Claude (v času pisanja).
2 – Poziv začnite z “Update your memory with the following:”.
3 – Najbolje je, da razložite, za katero situacijo bi želeli, da se zgodi ta “slog”.
Update your memory with the following:
Whenever I write an email with you, I never use the – em dash, ever.
Be straightforward. Don’t be too polite. Always add an actionable next step.
To je to. LLM bo v bodoče “vedel” vaše specifična želje in/ali okoliščine.
