Skip to content

Never stop coding. Free MIT AI gateway: one endpoint, 375 providers (150+ free), 1,400+ models Kimi, Claude, GPT, Gemini, GLM, DeepSeek, MiniMax. Works with Claude Code, Codex, Cursor, OpenCode, Cline & Copilot. Quota-aware auto-fallback, RTK+Caveman compression saves 15-95% tokens, MCP/A2A, Desktop/PWA. Built by 900+ contributors

48.8 KB
Raw Blame

🗜️ Prompt Compression Guide — OmniRoute (Hausa)

🌐 Languages: 🇺🇸 English · 🇪🇹 am · 🇸🇦 ar · 🇦🇿 az · 🇧🇬 bg · 🇧🇩 bn · 🇧🇦 bs · 🇨🇿 cs · 🇩🇰 da · 🇩🇪 de · 🇬🇷 el · 🇪🇸 es · 🇪🇪 et · 🇮🇷 fa · 🇫🇮 fi · 🇫🇷 fr · 🇮🇪 ga · 🇮🇳 gu · 🇮🇱 he · 🇮🇳 hi · 🇭🇷 hr · 🇭🇺 hu · 🇦🇲 hy · 🇮🇩 id · 🇳🇬 ig · 🇮🇹 it · 🇯🇵 ja · 🇬🇪 ka · 🇰🇭 km · 🇮🇳 kn · 🇰🇷 ko · 🇱🇹 lt · 🇱🇻 lv · 🇮🇳 ml · 🇮🇳 mr · 🇲🇾 ms · 🇲🇹 mt · 🇲🇲 my · 🇳🇵 ne · 🇳🇱 nl · 🇳🇴 no · 🇮🇳 or · 🇮🇳 pa · 🇵🇭 phi · 🇵🇱 pl · 🇵🇹 pt · 🇧🇷 pt-BR · 🇷🇴 ro · 🇷🇺 ru · 🇱🇰 si · 🇸🇰 sk · 🇸🇮 sl · 🇷🇸 sr · 🇸🇪 sv · 🇰🇪 sw · 🇮🇳 ta · 🇮🇳 te · 🇹🇭 th · 🇹🇷 tr · 🇺🇦 uk-UA · 🇵🇰 ur · 🇺🇿 uz · 🇻🇳 vi · 🇳🇬 yo · 🇨🇳 zh-CN · 🇹🇼 zh-TW


Ajiye 15-95% na mahallin da ya cancanta ta atomatik. Don taƙaitaccen bayani, duba sashen Matsewa na README.

Bayani Gabaɗaya

OmniRoute yana aiwatar da bututun matse umarni mai tsarin sassa wanda ke gudana tun da wuri kafin buƙatu su isa ga masu samarwa na sama. Wannan yana nufin ana samun tanadin token ɗinka ba tare da ka lura ba — babu buƙatar canza tsarin aikinka.

Buƙatar Abokin Ciniki
  → Mai Zaɓar Dabarar Matsewa
    → Akwai saitin combo na musamman? → Yi amfani da saitin combo
    → An kai ma'aunin farawa ta atomatik? → Yi amfani da yanayin auto
    → Yanayin tsoho? → Yi amfani da saitin gama-gari
    → A kashe? → Tsallake matsewa
  → Yanayin Matsewa da Aka Zaɓa
    → A kashe: Babu matsewa
    → Lite: Amintaccen tsaftace sarari/tsarawa (~15%)
    → Standard: Cire kalmomin cikewa irin na Caveman (~30%)
    → Aggressive: Tsufar tarihi + taƙaitawa (~50%)
    → Ultra: Tacewa bisa ƙa'ida + rage tubalan lamba (~75%)
    → RTK: Tacewar fitarwar terminal/kayan aiki mai fahimtar umarni (kewayon 60-90% a sama)
    → Stacked: Bututun injuna da yawa mai bin tsari, yawanci RTK sannan Caveman (kewayon cancanta 78-95%)
  → Buƙatar da Aka Matse → Mai Samarwa

Yanayin Matsewa

A Kashe

Ba a aiwatar da matsewa ba. Duk saƙonni suna wucewa ba tare da canji ba.

Yanayin Lite (~15% tanadi, jinkiri <1ms)

Yanayi mafi aminci — babu sauyin ma'ana, tsaftace tsarawa kawai:

Dabara Bayani
collapseWhitespace Haɗa jerin layukan da babu komai da sararin ƙarshen layi
dedupSystemPrompt Cire saƙonnin tsarin da aka maimaita
compressToolResults Matse dogayen fitarwar kayan aiki/ayyuka
removeRedundantContent Cire umarnin da aka maimaita
replaceImageUrls Gajarta URI na bayanan hoton base64

Ya fi dacewa da: Amfani na dindindin, hanyoyin aiki masu matuƙar buƙatar aminci.

Yanayin Standard (~30% tanadi)

An samo ra'ayinsa daga Caveman — yana cire kalmomin cikewa da dogayen salon magana tare da kiyaye ma'ana:

  • Yana cire kalmomin cikewa ("don Allah", "ina tsammani", "a taƙaice", "a zahiri")
  • Yana taƙaita dogayen jimloli ("domin a yi" → "don", "sakamakon" → "saboda")
  • Yana cire lafazin ladabi mai rangwame ("Za ka damu idan...", "Idan za ka iya don Allah...")
  • Dokokin regex sama da 30 da aka daidaita don umarnin rubuta lamba

Ya fi dacewa da: Ayyukan rubuta lamba na yau da kullum, ƙungiyoyi masu lura da farashi.

Yanayin Aggressive (~50% tanadi)

Sarrafa tarihi cikin hikima don zaman aiki masu tsawo:

  • Tsufar Saƙonni — ana ƙara matse tsofaffin saƙonni sannu a hankali
  • Matse Sakamakon Kayan Aiki — ana yanke ko ɓoye dogayen fitarwar kayan aiki (layukan farko/ƙarshe, tace layukan da suka dace, taƙaita maɓallan JSON)
  • Masu Kiyaye Ingancin Tsari — yana tabbatar da cewa nau'i-nau'in tool_use + tool_result sun kasance daidai
  • Fahimtar Tagar Mahalli — yana kiyaye iyakokin token na kowane samfuri

Ya fi dacewa da: Dogayen zaman gyaran matsala, manyan rumbunan lamba.

Yanayin Ultra (~75% tanadi)

Matsakaicin matsewa don yanayin da token yake da matuƙar muhimmanci:

  • Tacewa Bisa Ƙa'ida — tace token na rubutu bisa maki
  • Kiyaye Tsari — tubalan lamba masu shinge, lamba a cikin layi, URL da masu ganowa ana sanya musu alamar wucin gadi sannan a sake haɗa su yadda suke, ba a taɓa tace su ba
  • Matakin SLM na zaɓi — ƙaramin samfuri na cikin gida zai iya inganta tacewar idan an saita shi
  • Mai cin gashin kansa daga yanayin Aggressive: ba ya gudanar da tsufar saƙonni, matse sakamakon kayan aiki ko mai taƙaitawa na madadin (gazawar matakin SLM ce kaɗai za ta iya tura aikin madadin ta hanyar aggressive)

Ya fi dacewa da: Lokacin da kake kai iyakar mahalli akai-akai.

Yanayin RTK (kewayon 60-90% a sama)

An inganta yanayin RTK don dogayen fitarwar kayan aiki da ke bayyana a zaman wakilan rubuta lamba:

  • Yana gano nau'ikan umarni/fitarwa kamar git status, git diff, git log, masu gudanar da gwaji, ginawa ta TypeScript/Vite/Webpack, ESLint/Biome/Prettier, binciken npm/shigarwa, rajistan Docker, fitarwar infra, da fitarwar shell ta gama-gari
  • Yana amfani da tarin matatan JSON daga open-sse/services/compression/engines/rtk/filters/
  • Yana shigo da matatan RTK TOML schema v1 daga fayilolin filters.toml na aikin ko na gama-gari, tare da tabbatarwar gwaji a ciki da tantance amincewa ga fayilolin aikin
  • Ya ƙunshi matata 55 da aka gina a ciki tare da misalan tabbatarwa a ciki
  • Yana cire jerin sarrafa ANSI, sandunan ci gaba, layukan da aka maimaita, da hayaniyar da ba ta da amfani
  • Yana kiyaye gazawa, kurakurai, gargaɗi, fayilolin da aka canza, taƙaitawa, da ƙarshen doguwar fitarwa
  • Yana goyon bayan matatan aikin da aka tantance amincewarsu, matatan gama-gari, da dawo da ɗanyen fitarwa da aka ɓoye bayanansa a matsayin zaɓi

Ya fi dacewa da: Zaman wakilai masu ɗauke da shell, ginawa, gwaji, git, grep, da bayanan fitarwar fayil.

Yanayin Stacked (kewayon cancanta 78-95%)

Yanayin Stacked yana gudanar da injunan matsewa da yawa cikin ƙayyadadden tsari. Bututun tsoho shi ne:

RTK -> Caveman

Wannan tsarin yana fara taƙaita fitarwar terminal/kayan aiki, sannan ya yi amfani da taƙaitawar ma'ana ta Caveman ga sauran umarnin harshe na halitta. Ana iya saita bututun Stacked a matakin gama-gari ko ta hanyar haɗaɗɗun saitunan matsewa da aka sanya wa haɗaɗɗun saitunan turawa.

Ya fi dacewa da: Haɗaɗɗen mahalli mai manyan rajistan kayan aiki tare da umarnin ɗan Adam ko taƙaitawar mataimaki.


Lissafin Tanadin Upstream

OmniRoute yana bayyana tanadin matsawa daga tushe biyu: ma’aunin gwajin ayyukan upstream da kuma haɗin injunan OmniRoute da kansa.

Tushe Adadin README na upstream da aka yi amfani da shi a nan
Caveman ~75% raguwar token na fitarwa, matsakaicin tanadin fitarwa na gwaji 65%, kewayon 22-87%, da kayan aikin matsawar shigarwa na ~46%
RTK Tanadin fitarwar umarni na 60-90%; zaman misali na token ~118,000 -> ~23,900, ko tanadin 79.7% (~80%)

Ga bayanan kayan aiki/mahalli masu karo da juna, haɗin OmniRoute na asali yana jera injunan kamar haka:

RTK -> Caveman

Haɗaɗɗen tanadin na ninkawa ne, ba na tarawa ba:

haɗaɗɗe = 1 - (1 - tanadin RTK) * (1 - tanadin shigarwar Caveman)
matsakaici = 1 - (1 - 0.80) * (1 - 0.46) = 89.2%
kewayo     = 1 - (1 - 0.60..0.90) * (1 - 0.46) = 78.4-94.6%

Wannan adadin 78-95% yana aiki ne idan RTK da Caveman dukansu za su iya rage bayanan shigarwa/mahalli iri ɗaya. Yanayin fitarwar martanin Caveman daban yake: idan an kunna shi, yi amfani da tanadin fitarwa na Caveman (65% a matsakaici, ~75% a babban adadi, kewayon 22-87%). Jimillar tanadin kuɗin amfani ta dogara da haɗin prompt/fitarwar ku.

Abin da "cancanta" yake nufi a zahiri

Babban kewayon 15-95% gaskiya ne, amma yana aiki ne kawai ga bayanai masu maimaituwa ko yawan surutu — layukan kuskure masu maimaituwa, log na build da ke cika wuri da gargaɗi iri ɗaya, ko fitarwar grep/karanta-fayil mai girma fiye da kima. Wannan ba yana nufin kowace buƙata za ta tanadi adadin haka ba.

An tabbatar ta hanyar gwaji (tests/unit/compression/stacked-compression-tool-result-savings.test.ts): gudanarwar stacked (RTK + Caveman) a kan tubalin tool_result mai tsarin Anthropic wanda ya ƙunshi layukan kuskure iri ɗaya guda 300 ya samar da tanadin token na 95.93% / tanadin haruffa na 96.26% — daidai cikin kewayon da aka tallata. Amma idan aka gudanar da wannan pipeline ɗin a kan fitarwar kayan aiki ta yau da kullum, wadda ba ta da maimaituwa (jerin sakamakon grep mai tsabta, ɗan gajeren karatun fayil, rubutun tattaunawa na yau da kullum), daidai yake samar da tanadi kusa da sifili, saboda babu wani abu mai maimaituwa da za a cire, kuma validateCompression() (validation.ts) yana ƙin aika sake-rubutawar da za ta cire ko sauya tubalan lamba, URLs, kanun labarai, nau’ikan sigogi, ko masu gano constants na ALL-CAPS.

Wannan hali ne da ake tsammani kuma mai aminci, ba bug ba ne: zaman coding wanda galibi yake karantawa/grep na fayiloli masu tsabta zai ga ɗan ƙaramin jimillar tanadi ko da an kunna matsawa gaba ɗaya, yayin da zaman da ya ci karo da loop mai gazawa ko linter mai yawan surutu zai ga cikakken kewayon 78-95% a kan wannan zirga-zirgar. Kada ku ɗauki ƙaramin kason jimillar tanadin zama guda a matsayin hujjar cewa an saita matsawa ba daidai ba — da farko ku bincika ko ainihin fitarwar kayan aikin tana da maimaituwa.


Nuna Tanadin Token

Ba tare da matsawa ba: 47K token aka aika zuwa LLM
Tare da Lite:          40K token aka aika          (an tanadi 15% — aminci, kullum a kunne)
Tare da Standard:      33K token aka aika          (an tanadi 30% — ƙa'idodin caveman-speak)
Tare da Aggressive:    24K token aka aika          (an tanadi 50% — tsufar bayanai + taƙaitawa)
Tare da Ultra:         12K token aka aika          (an tanadi 75% — datsawar heuristic)
Tare da RTK:           19K-5K token aka aika       (an tanadi 60-90% a fitarwar umarni/kayan aiki)
Tare da Stacked:       10K-2.5K token aka aika     (kewayon RTK+Caveman mai cancanta na 78-95%)

Saitawa

Dashboard

Je zuwa Dashboard → Context & Cache:

  • Caveman — zaɓin yanayi, fakitin harsuna, samfoti, da tsoffin saitunan gama-gari
  • RTK — samfotoin tace umarni, saitunan tsaro na RTK, da kundin matata
  • Compression Combos — jerin injuna masu suna da aka sanya wa haɗaɗɗun turawa
  • Auto-Trigger Threshold — kunna matsawa ta atomatik idan adadin token ya wuce iyakar da aka saita

Keɓantaccen Saiti na Kowane Combo

A cikin Dashboard → Context & Cache → Compression Combos, sanya compression combo ga wani routing combo:

Combo: "free-tier-fallback"
  Compression Combo: "coding-agent-stack"
  Pipeline: RTK -> Caveman
  Targets:
    1. if/kimi-k2.7-code
    2. if/qwen3.8-max-preview

Wannan yana ba ka damar amfani da matsawa mai jeri a kan masu samar da sabis na kyauta/na coding, yayin da kake riƙe yanayin lite a kan biyan kuɗi masu kuɗi.

Wannan aikin sanya "Per-Combo Override" wani iko ne daban da keɓantaccen saitin yanayin matsawar routing-combo (Default/Off/Lite/Standard/Aggressive/Ultra/Codex Responses — tsarin bayanan filin yana kuma karɓar rtk, stacked da omniglyph) — wannan keɓantaccen saitin ba ya zaɓar wani jerin compression-combo mai suna; kawai yana saita filin compressionMode da resolveCompressionPlan ke dubawa. Ana iya saita shi ko dai a katin combo (Dashboard → Combos) ko kuma, tun daga #6760, ga kowane routing combo a cikin jerin "Assign to routing" da ke Dashboard → Context & Cache → Compression Combos, kusa da akwatin zaɓin sanya pipeline da aka bayyana a sama. Duk wuraren biyun suna adana canje-canje ta endpoint ɗin PUT /api/combos/{id} iri ɗaya.

Keɓantaccen saiti na kowace buƙata

Aika request header na x-omniroute-compression don sauya tsarin matsawa ga buƙata guda. Shi ne mafi fifiko — yana rinjayar keɓantaccen saitin routing-combo, profile mai aiki, auto-trigger, da Default na panel. Ana yin biris da ƙimomin da ba a sani ba (ba a taɓa ƙin buƙatar ba), kuma babban maɓallin gama-gari har yanzu shi ke sarrafa komai: idan an kashe matsawa gaba ɗaya, header ba zai iya kunna ta ba. Ƙimomi:

Ƙima Tasiri
off Babu matsawa ga wannan buƙatar.
default Default profile da aka samo daga panel (yana yin biris da active profile). Ana barin injunan lossy a kashe.
safe Daidai yake da rashin saka header: dedup da naɗe farin sarari kawai.
allow-lossy Riƙe tsarin operator na wannan buƙatar, gami da taƙaitawa, matatan dacewa, da sake rubuta salo.
engine:<id> Inji guda idan an kunna shi, misali engine:rtk. Wannan shi ne zaɓin kunna injin ga kowace buƙata.
<combo> Combo mai suna, ana fara daidaita shi ta suna (ba tare da la’akari da manya ko ƙananan haruffa ba), sannan ta id.

Ba tare da allow-lossy, engine:<id>, ko combo mai suna ba, ba a amfani da injunan lossy. Har yanzu buƙatar tana samun session dedup da naɗe farin sarari idan matsawa tana kunne.

Ana mayar da tsarin da aka yi amfani da shi a cikin response header na X-OmniRoute-Compression: <mode>; source=<source>, inda <source> yake ɗaya daga cikin request-header, routing-override, active-profile, auto-trigger, default, ko off.

API

# Samo saitunan matsawa
curl http://localhost:20128/api/settings/compression

# Sabunta saitunan matsawa
curl -X PUT http://localhost:20128/api/settings/compression \
  -H "Content-Type: application/json" \
  -d '{"defaultMode":"stacked","autoTriggerMode":"stacked","autoTriggerTokens":32000}'

# Yi samfotoin takamaiman payload na RTK/stacked
curl -X POST http://localhost:20128/api/compression/preview \
  -H "Content-Type: application/json" \
  -d '{"mode":"rtk","messages":[{"role":"tool","content":"npm test output here"}]}'

# Jera fakitin matatan RTK
curl http://localhost:20128/api/context/rtk/filters

# Gwada RTK kai tsaye tare da metadata na umarni na zaɓi
curl -X POST http://localhost:20128/api/context/rtk/test \
  -H "Content-Type: application/json" \
  -d '{"command":"npm test","text":"FAIL tests/example.test.ts\nError: boom"}'

Abin da Ake Karewa

Injin matsewa koyaushe yana kiyaye:

  • ✅ Tubalan lamba (masu shinge da na cikin layi)
  • ✅ URLs da hanyoyin fayil
  • ✅ Tsarukan JSON da bayanai masu tsari
  • ✅ Masu ganowa da alamomin fasaha da aka kare
  • ✅ Kalmomin lissafi
  • ✅ Ma'anar kiran kayan aiki/ayyuka
  • ✅ Umarnin tsarin (a yanayin lite)

Maido da ɗanyen fitarwa na RTK yana ɓoye maɓallan API na gama-gari, bearer tokens, Slack tokens, AWS access keys, kalmomin sirri, tokens, da sirrika kafin a adana komai.


Ƙididdigar Matsewa

Kowace buƙata da aka matse tana haɗa ƙididdiga a cikin rajistan uwar garken:

{
  "originalTokens": 47200,
  "compressedTokens": 40120,
  "savingsPercent": 15.0,
  "techniquesUsed": ["collapseWhitespace", "dedupSystemPrompt"],
  "mode": "lite",
  "engine": "caveman",
  "compressionComboId": "coding-agent-stack",
  "durationMs": 0.8,
  "rtkRawOutputPointers": []
}

Taswirar Matakai

Mataki Yanayi Matsayi
Mataki 1 Off, Lite ✅ An fitar
Mataki 2 Standard, Aggressive, Ultra ✅ An fitar
Mataki 3 RTK, Stacked, Compression Combos ✅ An fitar
Mataki 4 Output Styles, SLM-tier Ultra, tsarin gwajin kimantawa ✅ An fitar
Mataki 4C Kasafin mahallin da ke daidaitawa ("dial") — injin lissafi + API (contextBudget a kan PUT /api/settings/compression) + sarrafa yanayi/manufa na dashboard ✅ An fitar

Godiya

Dokokin matsewar yanayin Standard sun samo wahayi daga Caveman na JuliusBrussee (⭐ 51K+) — shahararren aikin "me ya sa a yi amfani da token da yawa alhali kaɗan za su yi aikin". Caveman ya bayar da rahoton raguwar tokens na fitarwa da ~75%, matsakaicin tanadin fitarwa na gwajin ma'auni da 65%, zangon fitarwa na 22-87%, da kayan aikin matse shigarwa na ~46%.

Yanayin RTK ya samo wahayi daga RTK - Rust Token Killer na RTK AI — aikin matse fitarwar umarni mai ƙarfi don tace fitarwar terminal, build, test, git, da kayan aiki. RTK ya bayar da rahoton tanadin 60-90%, inda samfurin zaman da ke cikin README ɗinsa ya nuna an adana ~80%.


Tsarukan Matsewa na Ci-gaba

Bayan hanyoyi 7 da aka bayyana a sama (tushen kuma yana karɓar hanyoyin codex-responses da omniglyph, waɗanda wannan jagorar ba ta rufe ba), sassan da ke ƙasa suna bayyana fasalolin da ke aiki a cikin ko tare da waɗannan hanyoyin: Matse Sakamakon Kayan Aiki da Tsufa Mai Ci-gaba su ne matakai na 1 da 2 na injin mai tsanani (yanayin Aggressive da matakin aggressive na bututun da aka jera), Bututun da Aka Jera shi ne yadda yanayin Stacked ke aiki, Matsewa Mai La'akari da Ma'ajiyar Wucin Gadi yana saukar da aggressive da ultra zuwa standard ga masu samarwa masu amfani da ma'ajiyar wucin gadi yayin da matsewa ke kunne, sannan Yanayin Fitowar Caveman da Salon Fitarwa umarnin faɗakarwar tsarin ne da ake zaɓar kunnawa, waɗanda suke kashe a tsoho, kuma suna tsara fitowar samfurin maimakon matse buƙatar.

Matsewa Mai La'akari da Ma'ajiyar Wucin Gadi

Wasu masu samarwa (kamar Anthropic mai amfani da ma'ajiyar wucin gadi ta faɗakarwa) suna goyan bayan ma'ajiyar wucin gadi ta faɗakarwa, wadda ke ba su damar adana wasu sassan faɗakarwar domin rage kuɗi da jinkiri. Lokacin da aka kunna ma'ajiyar wucin gadi, matsewa mai tsanani na iya lalata aiki saboda yana canza token ɗin da aka adana, yana mai kawar da ingancin ma'ajiyar.

Modul ɗin cachingAware.ts yana warware wannan ta hanyar gano mahallin ma'ajiyar wucin gadi da kuma daidaita dabarar matsewa yadda ya dace.

Yadda yake aiki

  1. Gano mahallin ma'ajiyar wucin gadi — Yana bincika jikin buƙatar don alamomin cache_control
  2. Gano masu samarwa masu amfani da ma'ajiyar wucin gadi — Yana bincika ko mai samarwar da aka nufa yana goyan bayan ma'ajiyar wucin gadi
  3. Daidaita dabara — Yana saukar da aggressive/ultra zuwa standard ga masu samarwa masu amfani da ma'ajiyar wucin gadi
  4. Tsallake faɗakarwar tsarin — Yawanci ana adana faɗakarwar tsarin, don haka kada a matse su

Mataimakin dabarar yana kuma dawo da tutar deterministicOnly, amma maginin tsari yana amfani da dabarar kawai — babu wani abu a matakan da ke biyo baya da ke karanta tutar a yanzu.

Misalin lamba

import {
  detectCachingContext,
  getCacheAwareStrategy,
} from "@omniroute/open-sse/services/compression/cachingAware";

const body = {
  model: "anthropic/claude-sonnet-4.5",
  messages: [{ role: "user", content: "Hello" }],
  cache_control: { type: "ephemeral" }, // ← Alamar ma'ajiyar wucin gadi
};

const ctx = detectCachingContext(body, { provider: "anthropic" });
// → { hasCacheControl: true, provider: "anthropic", targetFormat: null, isCachingProvider: true }

const strategy = getCacheAwareStrategy("aggressive", ctx);
// → { strategy: "standard", skipSystemPrompt: true, deterministicOnly: true }

Lokacin amfani

Matsewa mai la'akari da ma'ajiyar wucin gadi kullum yana kunne — ba a buƙatar wani saiti. Yana fara aiki duk lokacin da matsewa ke kunne kuma mai samarwar da aka nufa yana goyan bayan ma'ajiyar wucin gadi ta faɗakarwa (Anthropic, OpenAI, da sauransu); ba dole ba ne a samu alamomin cache_control a bayyane — kasancewar mai samarwa mai amfani da ma'ajiyar wucin gadi kaɗai yana jawo saukar da matakin, kuma alamomi kaɗai ba sa yin hakan (gano alama yana ba da bayanai ga na'urar auna ma'ajiyar wucin gadi, ba shawarar dabara ba).

Tsufa Mai Ci-gaba

Dogayen tattaunawa suna tara saƙonni masu yawa, amma tsofaffin saƙonni suna raguwa a muhimmanci. Modul ɗin progressiveAging.ts yana rage cikakken bayanin saƙonni bisa nisan zagaye (ana auna nisan daga ƙarshen tattaunawar). Tare da tsoffin saitunan da aka samar (verbatim: 2, light: 2, moderate: 3):

  • Zagaye 2 na ƙarshe (nisa ≤ 2): Ana adana su yadda suke
  • Nisa 3: Matsewar Caveman (cire kalmomin cike gurbi)
  • Nisa 4+: Ana taƙaita saƙonnin mataimaki; ana rage saƙonnin mai amfani zuwa layinsu na farko, tare da iyakar haruffa 120; ba a taɓa sauran rawar ba. Faɗakarwar tsarin, saƙonnin da aka riga aka tsufar, da sabon saƙon mai amfani koyaushe ana adana su yadda suke ba tare da la'akari da nisa ba. Ba a yar da komai gaba ɗaya, kuma ba za a iya isa rukunin light da tsoffin saitunan da aka samar ba (light yana daidai da verbatim).

Misalin lamba

import { applyAging } from "@omniroute/open-sse/services/compression/progressiveAging";

const messages = [
  { role: "system", content: "You are a helpful assistant" },
  { role: "user", content: "What is 2+2?" },
  { role: "assistant", content: "4" },
  // ... ƙarin zagaye 50 ...
];

const { messages: aged, saved } = applyAging(messages, {
  verbatim: 3, // zagaye 3 na ƙarshe: yadda suke
  light: 8, // nisa <= 8: matsewa mai sauƙi
  moderate: 20, // nisa <= 20: matsewar caveman
  fullSummary: 5, // nau'in yana buƙatarsa, amma lambar rarraba rukuni ba ta karanta shi
  // nisa > 20: ana taƙaitawa (mataimaki) / ana adana layin farko (mai amfani)
});

// saved = adadin token da aka adana

Lokacin amfani

Tsufa mai ci-gaba kullum yana kunne a yanayin aggressive — shi ne mataki na 2 na compressAggressive(). Yanayin Ultra ba ya gudanar da shi. Yana da tasiri musamman ga:

  • Dogayen zaman rubuta lamba
  • Tattaunawar da take ɗaukar kwanaki da yawa
  • Gudanawar aiki ta wakilai mai kiran kayan aiki da yawa

Yanayin Fitowar Caveman

Yanayin fitowar Caveman yana ƙara umarnin faɗakarwar tsarin waɗanda ke neman samfurin da kansa ya bayar da gajeriyar fitarwa — matakin lite yana neman amsoshi taƙaitattu waɗanda ke riƙe da cikakkun jimloli, full yana neman ya "amsa a taƙaice kamar caveman mai wayo", kuma ultra yana neman fitarwa irin ta saƙon gaggawa; umarnin roƙo kawai suke yi, ba za su iya tabbatar da hakan ba. Buƙatu suna karɓar su ta hanyar applyOutputStyles() (open-sse/services/compression/outputStyles/apply.ts): open-sse/handlers/chatCore.ts da farko yana warware zaɓin ta amfani da shim ɗin dacewa da tsohon tsari (resolveOutputStyleSelection() a cikin open-sse/services/compression/outputStyles/backCompat.ts), wanda, yayin da outputStyles ba ya ɗauke da komai, yake daidaita cavemanOutputMode da aka kunna zuwa salon fitarwa terse-prose a matakin cavemanOutputMode.intensity (duba Dacewa da tsohon tsari a ƙasa); ana amfani da zaɓin outputStyles marar komai ba yadda yake, sannan cavemanOutputMode.enabled da intensity ba sa yin wani tasiri, yayin da maɓallin autoClarity nasa har yanzu yake aiki. outputMode.ts yana ɗauke da rubutun umarnin (CAVEMAN_INSTRUCTION_BY_LANGUAGE), hanyar ƙetare abun ciki, da mataimakin sanya wurin da shigarwa ke amfani da shi; nasa mai shigarwa applyCavemanOutputMode() ba shi da wani mai kiransa a tsarin samarwa.

Yadda yake aiki

Wannan yanayin ba ya matse shigarwar. Yana ƙara toshe na umarni zuwa system prompt (duba Yadda sakawa ke aiki a ƙasa), kuma duk yanayin matse shigarwa da aka zaɓa don buƙatar har yanzu yana gudana daga baya, a kan jikin da yanzu ke ɗauke da toshen. Kafin jimlar iyakoki ta gama-gari da kowane mataki ke ƙarewa da ita, matakin full na Turanci yana cewa:

"Amsa a taƙaice kamar gogaggen mutumin kogo. Cire articles (a/an/the), kalmomin ciko (just/really/basically/actually/simply), gaisuwa, kalmomin nuna rashin tabbas. Yankakkun jimloli sun yi. Yi amfani da gajerun kalmomi masu ma’ana ɗaya (big ba extensive ba, fix ba implement ba). Kiyaye dukkan muhimman bayanan fasaha, code, errors, URLs, identifiers daidai."

Wannan yana aiki musamman da kyau don:

  • Samar da code (amsa mafi taƙaituwa = ƙarancin tokens)
  • Tambaya da amsa cikin sauri (ba a buƙatar dogayen bayanai)
  • Sarrafa tarin ayyuka (ƙara yawan abin da ake sarrafawa)

Lokacin amfani

Yanayin fitarwa na Caveman na buƙatar opt-in. Idan an kunna matsawa (enabled: true, babban maɓallin kunna a shafin Compression Settings), kunna shi da cavemanOutputMode.enabled; intensity yana zaɓar lite, full ko ultra:

{
  "enabled": true,
  "cavemanOutputMode": {
    "enabled": true,
    "intensity": "full"
  }
}

Maɓallin Output Mode na haɗin matsawa (outputMode, tare da mataki a outputModeIntensity) yana saita wannan maɓalli ga buƙatun da wannan haɗin ya shafa, kuma kayan aikin MCP na omniroute_set_compression_engine yana rubuta shi ta hanyar hujjar boolean ɗinsa ta outputMode. Zaɓin outputStyles marar fanko yana da fifiko a kan wannan maɓalli. A cikin dashboard, kunna salon fitarwa na Terse prose yana saka wannan toshe ɗin (duba Output Styles a ƙasa).

Salon Fitarwa (kasida)

Yanayin fitarwa na Caveman da ke sama shi ne tsohuwar hanyar salo guda ɗaya. Phase 4 ya faɗaɗa shi zuwa kasidar salon fitarwa masu iya haɗuwa: OUTPUT_STYLE_CATALOG a cikin open-sse/services/compression/outputStyles/catalog.ts. Kowane salo umarni ne na system-prompt wanda ke neman samfurin da kansa ya samar da fitarwa mai rahusa; ana iya kunna salo tare kuma ana saka su bisa jerin kasidar.

Salo id Abin da yake yi Harsunan umarni
Gajeriyar lafazi terse-prose Cire kalmomin ciko/articles/kalmomin nuna rashin tabbas; kiyaye muhimman bayanan fasaha daidai. Rubutu iri ɗaya da tsohon yanayin fitarwa na caveman (an yi nuni da shi, ba a sake rubuta shi ba). en, pt-BR, es, de, fr, it, ru, zh, ja, id, vi
Ƙarancin code less-code Tsarin YAGNI: mafi ƙarancin canjin da ke aiki, babu ƙirƙirar abstractions da ba a nema ba. en, pt-BR, es, de, fr, it, ru, zh, ja, id, vi
Ponytail (malalacin babban dev) ponytail "Mafi kyawun code shi ne code da ba a taɓa rubutawa ba": sake amfani > sake rubutawa, tushen matsala > alama, mafi gajeren diff mai aiki. en, pt-BR, es, de, fr, it, ru, zh, ja, id, vi
Ina da ADHD (aiki-farko) i-have-adhd Aiki da farko (command/path/snippet kafin bayani), matakai masu lamba da iyaka, mataki na gaba guda ƊAYA takamaimai, babu gabatarwa/taƙaitawa/kalmomin rufewa. An daidaita daga ayghri/i-have-adhd (MIT). en, pt-BR, es, de, fr, it, ru, zh, ja, id, vi
Matatar UI ta Antislop antislop-ui Dole kowace dabarar gani ta ci gwajin manufa (tsari, ainihi, sauƙin karantawa); babu gradients/glass/glow na tsoho, iyakantaccen palette, babu shimfidar template. Duk salon antislop biyu suna ƙara iyakar hana controls marasa aiki / ikirarin ƙarya. en, pt-BR, es, de, fr, it, ru, zh, ja, id, vi
Matatar rubutu da ɗan Adam ta Antislop antislop-copy Takamaiman rubutu (babu em dash, generic CTAs, buzzwords, rufewar chatbot, ƙagaggun bayanai) tare da dokokin mutane: bambancin launi na WCAG AA, focus mai gani, isa da keyboard, yanayin babu komai/loading/error, sake daidaitawa a mobile. en, pt-BR, es, de, fr, it, ru, zh, ja, id, vi
Taƙaitaccen CJK (文言) terse-cjk Amsa ta full/ultra cikin Sinanci na Gargajiya (文言); lite kawai yana neman gajerun amsoshi ba tare da function words, gaisuwa ko ƙawatawa ba. zh (an kayyade ta locale, duba ƙasa)

Kowane salo yana zuwa da matakan intensity uku — lite, full, ultra — kuma kowane mataki yana ƙarewa da jimlar iyakoki ta gama-gari (SHARED_BOUNDARIES a cikin outputMode.ts), wadda ke kiyaye code blocks, file paths, commands, errors da URLs daidai. Rubutun matakan terse-prose da terse-cjk suna ƙara identifiers cikin wannan jerin.

An kayyade terse-cjk ga locale na zh a wurare biyu. Shafin Compression Settings yana nuna layinsa ne kawai idan harshen UI na dashboard Sinanci ne (zh-CN ko zh-TW), kuma applyOutputStyles() yana saka shi ne kawai idan harshen da aka tantance na buƙatar (duba Zaɓin Harshe a ƙasa) shi ne zh. Ɓoye layin ba ya share zaɓin terse-cjk da aka adana: API na settings yana karɓar kowane style id, kuma adana wasu salo a shafin yana riƙe da shi. A lokacin buƙata, binciken harshe na applyOutputStyles() shi ne kaɗai ƙofar locale.

Yadda sakawa ke aiki

applyOutputStyles() (open-sse/services/compression/outputStyles/apply.ts) yana daidaita zaɓin da katalog (ana cire ids da ba a sani ba da salailan da harshensu bai dace ba, ba tare da kuskure ba; zaɓin da bai haifar da wani salo ba yana barin jikin ba tare da canji ba, ana tsallake shi a matsayin no_styles), yana haɗa umarnin da aka zaɓa bisa jerin katalog, yana ƙara ƙa'idar iyakoki sau ɗaya (tare da ƙa'idar tsaro, SAFETY_BOUNDARIES ko fassararta, idan an zaɓi less-code ko ponytail), kuma yana fara tubalin da alama guda ta hana maimaita aiki ([OmniRoute Output Styles]), don haka sake aiwatarwa ba ya yin komai. Idan harshen da aka tantance (duba Zaɓin harshe a ƙasa) yana da fassara, ana saka umarnin da aka fassara maimakon na Turanci.

A jikin da ke da jerin messages mara komai ba, ana gudanar da binciken hana maimaitawa kafin kaucewa abun ciki: idan alamar [OmniRoute Output Styles] ta riga ta kasance a filin system na babban mataki (string ko jerin tubalan abun ciki) ko kuma a saƙon system mai abun ciki na string, ana barin jikin ba tare da canji ba a matsayin already_applied kuma ba a gudanar da binciken kalmomin maɓalli. In ba haka ba, kaucewar abun ciki (shouldBypassCavemanOutputMode() a cikin open-sse/services/compression/outputMode.ts) tana duba rubutun saƙonni uku na ƙarshe, ko da menene matsayinsu, sannan ta tsallake salailan na dukan zagayen idan rubutun ya dace da kalmomin maɓalli na tsaro, aikin da ba za a iya juyawa ba, ko neman ƙarin bayani, ko wani jeri mai la'akari da oda: first, then, after that, before, rollback ko backup wanda delete, drop, migrate, deploy ko release ya biyo bayansa cikin haruffa 240. Kaucewar tana aiki yayin da maɓallin Auto-Clarity Bypass (cavemanOutputMode.autoClarity, wanda yake kunne ta tsohuwa) yake kunne; kashe maɓallin yana tsallake binciken kalmomin maɓalli.

Idan kaucewar ta ƙyale zagayen ya wuce, placeSystemInstruction() (a fayil ɗin nan), wanda ba ya taɓa ƙirƙirar sabon messages[0], yana sanya tubalin a na farko daga cikin waɗannan da ya samu:

  1. Saƙon system na farko mai abun ciki na string: ana ƙara tubalin bayan rubutunsa.
  2. Filin system na babban mataki: ana ƙara tubalin bayan rubutun string, ko a ƙara shi a matsayin sabon tubalin rubutu zuwa jerin tubalan abun ciki.
  3. Saƙon system na farko da ya zo daga baya mai abun ciki na string: ana ƙara tubalin bayan rubutunsa.
  4. Idan babu ɗaya daga sama: tubalin yana shiga sabon saƙon system a ƙarshen messages.

A jikin da ba shi da jerin messages (ko yake da wanda babu komai a ciki), ba a gudanar da kaucewar abun ciki kuma ba a duba filin system na babban mataki. Ana ƙara tubalin bayan rubutun filin instructions na string, sai dai idan filin ya riga ya ƙunshi alamar [OmniRoute Output Styles], inda a wannan yanayin ake barin jikin ba tare da canji ba a matsayin already_applied. Idan jikin ba shi da filin instructions na string amma yana ɗauke da input (string ko jeri), tubalin yana zama instructions, yana maye gurbin duk ƙimar da ba string ba da filin ya riƙe. Jikin da ba shi da filin instructions na string ko input na string ko jeri ana barinsa ba tare da canji ba kuma ana tsallake shi a matsayin no_messages.

Yadda ake kunnawa

A dashboard: Compression Context → Compression Settings (/dashboard/context/settings), sashen Output styles: layi ɗaya ga kowane salo tare da maɓallin kunna/kashewa da mai zaɓar mataki. Ana saka salailan yayin da compression kansa yake kunne (babban maɓallin shafin, enabled). Maɓallin Auto-Clarity Bypass yana kan shafin Caveman (/dashboard/context/caveman), a cikin katinsa na Output Mode. Ta hanyar shirye-shirye, tsarin compression yana adana zaɓin kamar haka:

{
  "outputStyles": [
    { "id": "i-have-adhd", "level": "full" },
    { "id": "less-code", "level": "lite" }
  ]
}

Daidaituwa da tsoffin nau'ikan: yayin da outputStyles yake babu komai, tsohon saitin cavemanOutputMode.enabled yana daidaituwa da terse-prose a cavemanOutputMode.intensity. Daga nan tubalin yana farawa da alamar [OmniRoute Output Styles], inda tsohon mai saka applyCavemanOutputMode() ya rubuta [OmniRoute Caveman Output Mode]. A ƙarƙashin alamar, rubutun yana daidai da tsohuwar sakawa a en, pt-BR, es, de, fr, it, ru, id da vi; a ja da zh yana ɗauke da ƙarin sarari guda kafin ƙa'idar iyakoki. Ana fassara terse-prose zuwa pt-BR, es, de, fr, it, ru, zh, ja, id da vi, don haka buƙatar da harshenta da aka tantance yake hu tana samun rubutun Turanci a inda tsohon mai saka yake amfani da na Hungarian.

Zaɓin harshen salo na fitarwa (resolveOutputStyleLanguage() a cikin outputStyles/apply.ts): idan languageConfig.enabled yana kunne, autoDetect yana ɗaukar samfurin sabon saƙon mai amfani na ƙarshe a cikin jerin messages na buƙatar wanda ke da rubutu (abun ciki na string, ko text na sassan abun cikinsa) sannan ya gudanar da mai ganowa na injin Caveman (detectCompressionLanguage()) a kansa. Mai ganowar yana mayar da zh ga rubutun da ke da haruffan Han kuma babu kana; in ba haka ba yana mayar da duk wanda yake da mafi yawan daidaituwar alamomi daga cikin it, pt-BR, es, de, fr, ru, ja, hu da id, sannan en idan babu wanda ya dace — rubutun da ba zai iya rarrabawa ba yana samun Turanci, ba defaultLanguage ba, kuma ba a taɓa gano vi duk da cewa salailan suna zuwa da rubutun vi. Jikin Responses API yana ajiye zagayensa a cikin input, wanda ba a ɗaukar samfurinsa, don haka yana samun defaultLanguage, sannan Turanci. Idan babu saƙon mai amfani a cikin messages da ke da rubutu, ko idan autoDetect yana kashe, ana amfani da defaultLanguage, sannan Turanci. Idan languageConfig.enabled yana kashe, harshen shi ne Turanci — sai dai idan haɗin compression ya shafi buƙatar (haɗin da aka sanya wa routing combo na buƙatar, ko tsohon haɗin compression da chatCore yake komawa gare shi don ginannen stacked pipeline): aiwatar da haɗi yana kunna languageConfig.enabled ga wannan buƙatar kuma yana saita defaultLanguage daga fakitin harsunan haɗin (ƙimar da aka adana idan tana ɗaya daga fakitin haɗin, in ba haka ba fakitin farko na haɗin, wanda ke komawa en ta tsohuwa), yayin da autoDetect da aka adana (wanda yake kunne ta tsohuwa) har yanzu yake aiki. Injin shigarwa na Caveman yana zaɓar harshen rule-pack ta wata hanya dabam — ga kowane ɓangaren rubutu, kuma idan auto-detect yana kashe, ana sarrafa shi ta enabledPacks.

An ƙayyade matrix ɗin salo × harshe ta tests/unit/compression/output-styles-i18n-matrix.test.ts: kowane salon katalog yana buƙatar shigarwa a cikin BASELINE_LANGUAGES na gwajin; salon da ba a taƙaita shi ga wani locale ba dole ne ya zo da fassarar pt-BR (an keɓe terse-cjk mai taƙaituwar locale daga wannan ƙa’ida) sai dai idan yana cikin KNOWN_ENGLISH_ONLY, wanda zai iya ƙunsar salo marasa kowane irin fassara kawai — salon da aka jera wanda yake da kowace irin fassara zai faɗi gwajin; haka kuma salo zai faɗi gwajin idan ya rasa harshen da shigarwar BASELINE_LANGUAGES ɗinsa ta jera. Don ƙara salo, duba EXTENDING_COMPRESSION.md.

Matse Sakamakon Kayan Aiki

compressToolResult() a cikin open-sse/services/compression/toolResultCompressor.ts yana matse rubutun sakamakon kayan aiki da dabaru 5. Yana gwada su a wannan jeri, kuma dabara ta farko da aka kunna wadda gwajinta ya dace da abun ciki ce ke ƙayyade sakamakon:

  1. fileContent: abun ciki mai layi 3 ko fiye wanda aƙalla layi ɗaya, bayan yin watsi da tazarar gabatarwa, yake farawa da import , export , function , class , const , let , var ko return (kalmar maɓalli tare da sarari), ko da if, for ko while sannan ( ko (, yana riƙe layuka 20 na farko da 5 na ƙarshe, tare da alamar nuna an cire tsakiyar.
  2. grepSearch: abun ciki mai aƙalla layi ɗaya mai tsarin <path>:<digits>:, inda rubutun da ke gaban colon na farko ba shi da whitespace, yana riƙe waɗannan layuka kawai, har zuwa 30, sannan ya biyo da adadin duk ƙarin abubuwan da suka dace da kuma jerin fayilolin da suka dace; ana zubar da kowane sauran layi. Layi guda irin wannan ya isa ya kunna dabarar, don haka layin log da ya fara da timestamp kamar 12:30:45 shi ma yana ƙirguwa.
  3. shellOutput: fitarwa da ta ƙunshi jerin ANSI CSI (ESC[ sannan lambobi ko semicolons sannan harafi, kamar a lambobin launi) ko $ da whitespace ya biyo bayansa a ko’ina cikin rubutun tana rasa waɗannan jerin (ana riƙe sauran escapes, kamar ESC[?25l ko jerin taken taga na OSC) kuma tana riƙe layuka 50 na ƙarshe, tare da haɗe maimaitattun layuka masu jere. Saboda wannan gwajin yana gudana kafin json da errorMessage, fitarwar JSON ko ta kuskure da ta ƙunshi irin wannan $ ba za ta taɓa isa gare su ba muddin shellOutput yana kunne.
  4. json: JSON payload mai haruffa sama da 2,000 wanda yake farawa da { ko [ (bayan whitespace na zaɓi) kuma aka iya parse ɗinsa ana taƙaita shi: array mai abubuwa fiye da 7 yana riƙe abubuwa 5 na farko da 2 na ƙarshe da jimillar adadinsa, yayin da object yake riƙe keys 20 na farko, tare da maye gurbin kowace nested object ko array value da placeholder na {…N keys} (ga array, N shi ne tsawonsa) da alamar _remaining_<N>_keys mai ƙirga keys da aka zubar bayan 20 na farko. Ana kwafe scalar values gaba ɗaya, don haka object mai keys 20 ko ƙasa da haka wanda ba shi da nested values kawai ana sake tsara indentation ɗinsa — minified object yana ƙaruwa da haruffa kuma ya kasance ba a canza shi ba.
  5. errorMessage: fitarwa da ta ƙunshi, a ko’ina kuma ba tare da la’akari da girman haruffa ba, error:, error (kalmar da sarari ya biyo baya, kamar a no error found), [error], exception:, exception , [exception] ko traceback tana riƙe layinta na farko, layuka 10 na gaba da 3 na ƙarshe, tare da alamar … [N frames elided] … a madadin layukan da ke tsakaninsu. Alamar tana bayyana ne kawai idan layuka fiye da 13 suka biyo bayan layin farko, don haka fitarwar kuskure mai layuka 14 ko ƙasa ba a gajarta ta (a layuka 12 ko 13, layuka 3 na ƙarshe suna maimaita layukan da aka riga aka riƙe).

Bayan dabara ta dace, ko da wadda ba ta adana komai ba, ba a gwada dabarun da ke baya. Idan dabarar da ta dace ba ta adana estimated tokens ba (tsawo ÷ 4, an zagaye zuwa sama) — misali fayil mai kama da code mai layuka 25 ko ƙasa, ko JSON array mai haruffa sama da 2,000 wanda yake da abubuwa 7 ko ƙasa — aggressive engine yana riƙe ainihin sakamakon kayan aikin: duk masu kira (compressAggressive() da compressAnthropicToolResultBlock()) suna riƙe ainihin idan saved ya zama 0 ko ƙasa, yayin da compressToolResult() da kansa har yanzu yake mayar da fitarwar dabarar. Matakin sakamakon kayan aiki ba shi ne hukunci na ƙarshe ba: fallback summarizer na engine har yanzu zai iya gajarta saƙon tool ko function mai tsawon sama da haruffa 8,192 (maxTokensPerMessage, 2,048, sau 4).

Lokacin amfani

Matse sakamakon kayan aiki shi ne mataki na 1 na aggressive engine (compressAggressive() a cikin open-sse/services/compression/aggressive.ts), don haka yana gudana a yanayin Aggressive da kuma a matakin aggressive na stacked pipeline. Yana matse saƙonnin tool da function masu tsarin OpenAI da kuma rubutun da ke cikin blocks na Anthropic tool_result. Kowace dabara tana da nata switch ƙarƙashin aggressive.toolStrategies, duk suna kunne ta default. A dashboard, switches ɗin suna cikin kallon Advanced na shafin Caveman yayin da compression yake kunne kuma default mode ɗin shi ne Aggressive.

Stacked Pipeline

Yanayin stacked yana gudanar da engines da yawa a jere — yawanci RTK da farko (ajiya 60-90% a kan fitarwar kayan aiki), sannan Caveman a kan sauran rubutun (~46% na ajiyar input). Idan aka haɗa, wannan shi ne eligible range na 78-95% (duba Upstream Savings Math a sama): 1 - (1 - 0.60..0.90) × (1 - 0.46) matsakaicinsa ≈89%.

Yadda yake aiki

Input (tokens 1000)
  → RTK (tace mai fahimtar command) → tokens 200
    → Caveman (cire ƙarin kalmomi) → tokens 108
  → Output (tokens 108, ~89% na ajiya)

Lokacin amfani

Yi amfani da yanayin stacked don:

  • Workflows masu yawan amfani da kayan aiki (agentic coding, bincike)
  • Batch processing mai kula da tsada
  • Lokacin da kake buƙatar matsakaicin ajiyar token

Ana saita stacked pipelines ta global stackedPipeline compression setting, ko ta named compression combo da aka sanya wa routing combo (duba Per-Combo Override a sama) — ba ta auto-combo modePack ba (wannan field ɗin yana sake daidaita nauyin zaɓin model na auto-combo kawai, kuma stacked ba ingantaccen sunan pack ba ne).


Sauya Saitunan Haɗin Matsewa

Za ka iya sauya yanayin matsewa na gaba ɗaya ga kowane haɗi don daidaita halayya ga lokutan amfani daban-daban:

{
  "id": "coding-combo",
  "strategy": "priority",
  "config": {
    "weights": { "taskFit": 0.5 },
    "modePack": "quality-first"
  },
  "compressionOverride": "aggressive"
}

Wannan yana da amfani ga:

  • Haɗin rubuta lamba: Yi amfani da yanayin aggressive don zaman aiki masu tsawo
  • Haɗin tambaya da amsa cikin sauri: Yi amfani da yanayin lite don amsoshi masu sauri
  • Haɗin da suka dogara sosai da kayan aiki: Yi amfani da yanayin stacked don samun mafi girman tanadi
  • Haɗin samarwa: bar sauyin saitin a kashe ga masu samar da caching — daidaitawar da ke la'akari da cache wadda take aiki koyaushe tana saukar da aggressive/ultra zuwa standard kai tsaye (babu yanayin cache-aware da za a iya zaɓa)

Duba Kuma