Lesson 142 — Advanced

Deep Research & Long-Horizon Scenarios

Kuch sawaal ek prompt aur ek response ke andar hal ho jaate hain; kuch ghanton chalte hain, daurlag sources padhte hain aur ek report par khatam hote hain. Is lesson mein hum dekhenge long-horizon run alag kyun hota hai, question ko kaise decompose karte hain, breadth-first padhai, sources ko critically padhna, disagreement reconcile karna, notes carry-forward, stopping conditions aur audit-wali report.

Long-horizon run kya hai

Funda ye hai ki context window poore run tak nahi chalta — sources itne hote hain ki raw text carry karna mumkin hi nahi. Isliye is game mein jeet uski hai ki tum raw source ko tod kar chhote notes banao, notes ko aage carry karo, aur original sawaal ko kabhi bhoolo na.

Decompose karo, breadth-first jao

# Pehle sawaal ko chhote, independently checkable hisson mein todo
question = "Kya hum is vendor ko production ke liye lein?"
sub_qs = [
    "vendor kya support karta hai (SLA, uptime)?",
    "pricing model kya hai, hidden limits?",
    "hamari data residency requirement match hoti hai?",
    "alternatives kaun hain aur unka trade kya hai?",
]
# har sub-question ka apna sources + finding hoga — report isi pe baithegi

# Breadth first: pehle bahut candidates, phir chhun ke gehrai
candidates = search_all(sub_qs)              # kaafi saare links (titles/summaries)
worth = rank_and_pick(candidates, k=5)       # sirf chhune hue 5 ko close-read karo

• Question abhi bhi vague hai toh decompose mat karo — pehle scope, timeframe aur audience clear karo, warna sub-questions bhi utne hi fuzzy honge.
• Breadth-first ka logic simple hai: gehrai mehngi hai. 20 links ka title/summary scan karke 5 chunna, 20 ko gehrai se padhne se sasta bhi hai aur behtar coverage bhi deta hai.
• Har sub-question independent chal sakta hai — isse parallelism aata hai, aur ek ka rukna doosre ko nahi rokta.
• Jis sub-question ka jawab mil gaya use close karo — open questions ka chhota list hi tumhara run state hai.

Critical reading aur disagreement

• Source ka type pehchano: primary evidence (official docs, raw data), secondary commentary, marketing material, outdated claim — har claim ko uske type ke hisaab se weight do.
• Marketing pages facts bhi deti hain aur framing bhi — feature list verify karne layak hai, "best" wale claims nahi.
• Sources disagree karein toh average mat karo: behtar evidence prefer karo, warna disagreement ko explicitly note karo — "A yeh kehta hai, B yeh, evidence B ki taraf jhukti hai" ya "genuine uncertainty".
• Dates check karo — pricing, APIs aur terms badalte rehte hain; purana source naya fact nahi ban jaata.

Critical reading ka matlab har sentence pe doubt nahi — evidence ke tier ko pehchanna hai. Ek official changelog ka weight tweet se zyada hai, aur dono ka weight ek controlled observation se kam. Ye hierarchy tumhare notes mein visible honi chahiye, taaki synthesis mein hierarchy bik jaaye nahi.

Notes aur carry-forward state

# Context badh raha hai — raw sources mat carry karo, findings carry karo
note = {
    "sub_q": "pricing model?",
    "finding": "per-seat + usage tiers; free tier limits documented",
    "sources": ["doc:pricing", "doc:limits"],   # har finding ke sources
    "confidence": "verified-2-sources",
    "checked_at": "2026-10-19",
}
run_state = [note, ...]        # chhota structured state = tumhari memory

# Agla step: notes padho -> gap dhoondo -> next search chalao
# Raw page text sirf tabhi dobara khulta hai jab note kam pade

• Interim summary har kuch steps pe likho: kya mila, kya nahi mila, agla kya. Ye tumhara crash-recovery point hai — run toote toh isi se resume.
• Carry-forward object chhota rakho: finding + sources + confidence. Raw excerpts zaroorat pe re-fetch ho sakte hain, prompt mein baithe rehne ki zaroorat nahi.
• Har finding ke saath sources attach rakho — isi se report mein citation aayegi aur reader bina dobara search kiye verify kar sakega.
• Original question aur open sub-questions ko state mein pin karo — chhota sa ye cheez hi drift ko rokti hai.

Stopping, report aur classic failures

• Stopping conditions teen hain: (1) jawab mil gaya, (2) sources repeat hone lagein — naya kuch nahi aa raha, (3) baaki uncertainty search se resolve hi nahi hogi. Inke alawa rukna premature hai.
• Auditable report ka structure: question, short answer, evidence per claim, contradictions jo mile, gaps jo reh gaye — har claim ke paas source.
• Long-run ki classic failures: original question se drift, thin evidence pe confident synthesis, aur aisi citations jo sentence ko support hi nahi karti.
• Citation check har line pe karo: agar source wo nahi kehta jo line kehti hai, wo hallucinated synthesis hai — report ka sabse khatarnak bug.
• Gaps ko chhupao mat — "ye verify nahi ho paya" reader ke liye useful hai, chuppi nahi.

Exercise

Common mistakes

• Pehle hi source pe gehrai shuru kar dena: Pehla source hi poori picture bana deta hai — counter-evidence aur alternatives reh jaate hain. Fix: pehle breadth scan + shortlist, sirf top ko close-read karo.
• Disagreeing sources ka average nikaalna: Average kisi ne bola nahi hota — ek fabricated middle paida ho jaata hai. Fix: evidence hierarchy lagao, behtar wala prefer karo, warna disagreement likh do.
• Structured notes ke bina raw text carry karte rehna: Context exhaust hota hai, purane sources dobara padhne padte hain, resume bhi nahi hota. Fix: har step pe finding + sources + confidence wala chhota note.
• Thin evidence pe confident report likh dena: Reader ko lagta sab verified hai aur ek citation gir gayi toh pura trust khatam. Fix: confidence flag aur gaps section — jahan evidence nahi, wahan explicitly likho.

sawaal -> kaagazat list -> folders scan -> zaroori kholo -> har finding pe page number
      -> summary: kya mila, kya nahi, kya doubt hai      (cited notes)
warna: bina reference wali rai = kisi kaam ki nahi