Lesson 76 — Intermediate

Compression & Compaction

Context badhta jaata hai, budget fix hai — isliye beech ka rasta hai compaction: size kam karo par task-critical information bachao. Is lesson mein history summarize, tool outputs slim aur facts pin karna seekhoge.

Reduction vs information loss

Compression ka golden rule: shabd bacha sakte ho, facts nahi. Ek number, ID ya decision gayab ho gaya toh agent dobara kaam karega — cost aur confusion dono badhenge.

History compress karo

Nayi summary mein kya preserve karna hai: goal (verbatim), constraints, decisions + kyun, unresolved work. Baaki turns ka narrative ghulao, last kuch turns raw chhodo.

def compact_history(turns, summary, keep_last=4):
    if len(turns) <= keep_last:
        return summary, turns
    new_summary = summarize(
        "Purani summary:\n" + summary +
        "\nNayi conversation — facts, decisions, open items only:\n" +
        render(turns[:-keep_last]))
    return new_summary, turns[-keep_last:]
# Goal hamesha VERBATIM pin rehta hai — summarize nahi hota

Tool outputs slim karo

• Fields slice: model ko sirf wahi fields do jo usne maangi — bade API dumps nahi.
• Paginate: pehla page + cursor; details chahiye toh agla call (full result external store mein, L6.6).
• Dedupe/sort: repeat rows hatao, top-k rank karo, relevance se chhota karo.
• Errors chhote: status + short message; poora stack trace model ko nahi, log mein.

def slim(result, keep=5):
    return {
        "total": len(result["items"]),
        "items": result["items"][:keep],
        "next_cursor": result["next"],   # details phir se maang lenge
    }

Rolling & hierarchical summaries

Lambi chalne wale tasks mein ek summary kaafi nahi — layers banao: turn summary → session summary → project summary. Har layer mein upar wali ka compressed version + naya content.

• Rolling: har trigger par old turns summary mein ghul jaate hain, recent raw rehte hain.
• Hierarchical: hafte/chapters ke summaries alag — bade archives mein sab kuch ek summary mein mat gholo.
• Pinned facts alag: IDs, URLs, exact values, open TODOs — kabhi summarize nahi, sirf copy hote hain.
• Verbatim decisions: "schema X choose kiya (kyun)" — narrative compress, decision line pin.

Failure modes

• Goal drift: summary mein goal badal gaya — agent alag kaam karne lagta hai (goal verbatim pin karo).
• Over-compression: instructions bhi ghul gayin — constraints hamesha full text.
• Double compression: bina caps ke har trigger par chhota hota jaana — hierarchical + min-size cap rakho.
• Fabricated summary: summarizer ne fact jod diye — summary generation bhi structured output se validate karo.

Exercise

Common mistakes

• Goal bhi summarize kar dena: Goal ghula toh agent ka compass gayab. Goal + constraints ekdum raw pinned block mein — compression se exempt.
• Tool dump seedha andar: Ek paginated API ka full dump window ka aadha kha jaata hai. Slim preprocessor mandatory hai har tool result par.
• Summary of summary of summary: Compounding loss se facts phisalte jate hain. Hierarchical layers + verbatim anchors rakho, min-size cap ke saath.

summary = "Data fetch done, schema X decide (kyun: ...)"
pinned  = {"ids": [...], "numbers": {...}, "todo": [...]}
recent  = last_4_turns_raw   # dono ke upar recent raw