SearchOS
Intelligent Search Framework
$

SearchOS

A research workspace for the Web and your documents. Coordinate agents, follow original sources, and turn research methods into reusable skills.

Read Paper GitHub Local Search Demo Get Started
Overview

Project Overview

Search Reasoning State Management Evidence Extraction Skill Reuse Error Oversight Coverage-Driven

SearchOS maintains tasks, evidence, coverage, and research methods as persistent system state. Agents can continue from known facts, inspect unresolved questions, and reuse source access methods across Web and local research.

Web and Hybrid research use an Orchestrator to plan and dispatch work against SOCM snapshots. Local questions use a document QA agent directly. Composable middleware manages context, evidence intake, budgets, and progress around agent execution.

Problem Formulation · Open-Domain Information-Seeking Task

A query q decomposes into a chain of sub-question nodes; each table Tm (an entity×attribute coverage grid) frames the nodes it grounds. Two shapes recur: tables partition the chain to cover breadth, or tables overlap so that the node(s) framed by all of them form the answer set satisfying every constraint.

(a) BREADTH — tables partition the chain Query q decompose → sub-questions T₁ T₂ T₃ q₁ q₂ q₃ q₄ q₅ q₆ Answer = ⋃ₘ Tₘ comprehensive coverage report · union of regions Tables split the sub-questions into disjoint regions → broad coverage. e.g. “survey / compare all items in a field” (b) CONSTRAINT — nodes framed by all tables Query q “the one X with φ₁ ∧ φ₂ ∧ φ₃” T₁·φ₁ T₂·φ₂ T₃·φ₃ e₁ e₂ The node(s) framed by every table form the answer set: E* = T₁ ∩ T₂ ∩ T₃ → all entities satisfying φ₁ ∧ φ₂ ∧ φ₃ candidate entity / sub-question answer entity (e ∈ E*) grounded cell
(a) tables partition the chain — the answer is the union of covered regions (breadth). (b) tables overlap — the node(s) inside every table form the answer set E* (the intersection).

Agent Lifecycle · Multi-Agent Scheduling Timeline

A worked case: q = “Compare leading 2025 LLMs — context window, price, benchmark”. Swimlanes per agent, time flows left→right. Solid arrows = dispatch / trigger (control); dashed arrows = return result.

Scope Dispatch Monitor Dispatch+ Synthesize Orchestrator Explore Search Agents pipeline-parallel Writer long-horizon ReAct controller (always alive) create schema scope entities & sources Agent 1 · Task A search open find Agent 2 · Task B search open find Agent 3 · Task C search open find freed slot reused immediately search open find Task D complete chains overlap in time → Agent pipeline parallelism price / gaps (backfill) WriterTrigger sensor comparison table → draft → refine (grows with evidence) time dispatch / trigger return result
One run of the case query. The Orchestrator — the only long-lived agent — creates the schema, dispatches waves of short-lived sub-agents, and converges. The search agents run pipeline-parallel: their search → open → find stages overlap across agents and return asynchronously, so a freed slot is reused the moment a task finishes — a backfill wave later closes any gaps, and a coverage sensor auto-triggers the Writer partway through. Wall-clock tracks the slowest chain, not the sum, collapsing a long serial crawl into a few overlapping waves.

Core Capabilities

  • Automatic schema extraction (entity × attribute)
  • Coverage-driven exploration priority
  • Evidence Graph auto-extraction
  • Trace-driven skill creation and updates

Research Agents

  • explore_agent — Disambiguation / candidate enumeration
  • search_agent — Web or local evidence collection
  • writer_agent — Optional structured drafting

The Orchestrator normally produces the final answer. Local mode uses a dedicated document QA flow.

Quick Start

./install.sh
source .venv/bin/activate
python -m pip install -e ".[local-docs]"
searchos --setup
./web/start.sh
Updates

Feature updates

  1. Live Steering & Follow-ups

    Follow-ups now wake a waiting Orchestrator immediately, so it can act on your request while sub-agents are still running.

    How live steering works

  2. Portable Skills

    Create and update skills from tasks and traces, then export SKILL.md packages and CLI scripts for reuse in compatible harnesses.

    Skill creation and export

  3. Composable Middleware

    Context management, evidence extraction, and budget and progress checks can now be assembled as independent components for each agent.

    Component guide

  4. Knowledge Library

    Organize documents and saved Web sources by topic, browse them in list or graph views, and choose files or a whole library for research.

    Knowledge library guide

  5. Multimodal Search

    Source reading extends to images and scanned documents, with citations tied to the relevant page or region.

    Document parsing and citations

  6. Local & Hybrid Search

    Hybrid builds on Local Search to combine document reading and Web retrieval in one research workflow, with an explicit local source scope.

    Watch Local Search
Web & TUI introduction

Parallel research, shared evidence, and structured answers.

Product Tour

See SearchOS in Action

Follow the Web research workflow: connect providers, build a schema, inspect evidence, and export results. Available in English and Chinese.

Configure providers and API keys Build structured, evidence-backed research Repair gaps and export final reports
Architecture

System Architecture

How SearchOS’s five layers couple and collaborate — control, data, state, signal, and skill flow.

Relationship View · How Five Layers Collaborate

Click a legend button to isolate a flow type; hover a component to trace its connections.

SearchOS · Open-Domain Information Seeking Framework control · data · state · signal · skill flows across five layers — read left to right INPUT CONTROL EXECUTION GOVERNANCE MEMORY / STATE signal · force converge read SOCM snapshot — the Orchestrator decides on live state task-relevant state to agents User Question q natural language input Orchestrator long-horizon controller explore · schema · enqueue check · synthesize the sole decision-maker Agent Teams Explore Agentscope · disambiguate · discover Search Agent× N running in parallel search open find Writer (optional) short-horizon · clean local context Middleware · kernel Context LayerSOCM projection · history · trim Sensor Layerbudget · repetition · progress Extraction Layerbuffer pool · LLM extract · SOCM writer SOCM · shared state 1 · Frontiertyped priority queue explore τ1 · p=0.92 search τ2 · p=0.78 verify τ3 · p=0.63 2 · Evidence Graphclaims + provenance SourceA Claimcell support / conflict / refine 3 · CoverageMapmulti-table fill state EntityRankEvidStat filled · uncertain · missing 4 · Failure Memorydistilled dead-ends · shared repeated query search-without-open low-yield source ↺ shared by every search agent Skill Library Access · Strategy · Orchestrator Tools / Web + Local browser · local reader · skill CLI dispatch ‖ results write inject / block observation search · open · find select task-relevant skill context explore creates the coverage table shared failure memory — read by every parallel search agent Iterative Loop 1 · schedule from Frontier → 2 · agents explore (parallel) → 3 · middleware extracts → 4 · SOCM updates → 5 · synthesis & converge SearchOS · Open-Domain Information Seeking Framework sensor intervention · converge / replan live SOCM snapshot task-relevant state search · open · find select inject task-relevant playbook INPUTCONTROLAGENT TEAMRUNTIMESHARED STATE User Question q open-domain task Orchestrator long-horizon · sole decision-maker create relational schema schedule · monitor · backfill assess · synthesize · converge Specialized Agents Explore Agentscope · disambiguate · discover entities Search Agents × Npipeline-parallel evidence collection search open find Writer (optional)evidence-grounded report synthesis Search Tool Middleware Harness ContextSOCM projection · history · trim Sensorloop · stall · budget · intervention Evidence Extractionextract · anchor citation · write SOCM Tools / Web + Local browser · local reader · skill CLI SOCM · Shared Search State Frontierpriority task DAGpending · runningblocked · completed Evidence Graphclaim + provenance Coverage Maprelational schema state Failure Memoryshared dead endsqueries · sourcesskills · branches Portable Skill Library discover · load · execute · export ITERATIVE LOOP schedule→ agents run pipeline-parallel→ middleware extracts→ SOCM updates→ coverage-driven backfill→ synthesize
Tip — Click a legend button to filter a relationship type, or hover over a component to highlight its connections.
Architecture

Architecture Design

Three parts make the diagram work: the agents that decide, the state they share, and the middleware that runs around every step.

1

Multi-Agent System

For Web and Hybrid research, the Orchestrator plans work, dispatches source-specific tasks, and produces the final answer from shared evidence. Local mode uses a document QA agent directly.

Explore Agent the scout

When a question is vague, it does a quick reconnaissance pass — what's out there, which entities exist, where the ambiguities are — so the Orchestrator can draft a sensible plan. It writes no answer.

Optional reconnaissance uses search and reading tools to establish scope, identify entities, and resolve ambiguity before focused evidence collection.

Search Agent the investigator

Given a research question or missing facts, each agent searches and reads within its assigned source scope. Web and local agents can work in parallel on independent parts of a Hybrid task.

Web tasks use browser tools and loaded Access packages through exec. Local tasks use local_catalog, local_find, and local_open. Source middleware passes actual reading windows to evidence extraction.

Writer Agent the author

When enabled, the Writer uses collected evidence to draft a report section by section, with source citations. It is an optional delivery path; the Orchestrator normally writes the final answer.

The optional Writer uses outline and SOCM reading tools, including write_section, read_coverage, and read_evidence. Its tool scope is drafting from collected evidence.

The Orchestrator coordinates the research — it reads current state and decides what to dispatch, continue, or stop. Each sub-agent has its own execution context while committed findings remain available through shared state. Tap a card for the real tools & triggers.

2

State Model

The shared memory. Everything the agents learn is folded into four living objects — not a growing chat log.

Frontier the to-do list

A priority queue of search tasks. Each task moves pending → running → done (or blocked / cancelled), so the system always knows what's left to explore.

FrontierMemory of FrontierTask. Orchestrator pushes via enqueue_tasks; the Scheduler dispatches, unblocks, and reaps.

Evidence Graph the facts

Committed findings retain their sources and supporting passages. Evidence can come from Web pages, local documents, or saved skill results; competing values are checked before a conflict is confirmed.

EvidenceGraph. Source-specific middleware submits reading windows through ExtractionCoordinator; evidence retains its provenance.

CoverageMap the progress sheet

Entity × attribute tables track which questions have evidence and which need further research. New findings extend the tables without requiring agents to reconstruct them from conversation history.

CoverageMap of CoverageCell (CellStatus). Seeded by create_schema(), backfilled as evidence grows; drives gap detection and Writer structure.

Strategy Log the lessons

A record of what isn't working — repeated queries, dead-end sources — so the system stops digging in the same dry hole.

StrategyMemory (state attr strategy_log): StrategyPattern anti-patterns + FailureMemory. Feeds the Sensor layer and post-mortem distillation.

All four are persisted to searchos_workspace/<session>/search_state.json with atomic file locking, so multiple sub-agents can update them in parallel without clobbering each other. Tap a card for the real class names.

3

Middleware

Context, Sensor, and Extraction components are selected explicitly for each agent. Default pipelines supply common combinations; callers can assemble their own middleware lists.

Context prepare the next model call

Projects current research state into the prompt, manages reading history, and trims model input. Optional context compaction saves a continuation summary in checkpointed state.

SOCMContextMiddleware · SearchEpisodeMiddleware · DynamicTrimMiddleware · optional CompactionContextMiddleware.

Sensor after a tool returns

Tracks tool budgets, repeated queries, reading activity, and research progress. Components issue reminders or stop signals; the agent runtime applies the resulting controls.

BudgetMiddleware · QueryRepetitionMiddleware · ProgressGuardMiddleware · CoverageStallMiddleware. Mutable state is scoped to each agent.

Extraction turning pages into facts

Turns actual reading windows into source-backed findings. Web, local, and final-summary intake can be selected separately, sharing an extraction coordinator when needed.

WebEvidenceMiddleware · LocalEvidenceMiddleware · optional FinalSummaryEvidenceMiddleware. Explicit extraction policies select task rows or schema-routed facts.

Components share an explicit lifecycle. The runtime waits for triggered extraction before the next model step and settles remaining work on completion, failure, or cancellation. Model compatibility, telemetry, and execution control remain runtime responsibilities. Tap a card for the real class names.

Hook Timing View · How a Sub Agent Round Is Wrapped

Model and tool hooks apply the selected components. Source intake sends reading windows to the coordinator; the runtime owns synchronization and finalization.

Agent Loop — time flows right Before model_callContext projection / trim LLM DecisionProduces tool_call or stop message Before tool_callHarness budget / hard-block Tool Executionsearch / open / find After tool_callSensor + Extraction awrap_model_call1. Wait for triggered extraction2. Project SOCM and pending signals3. Prepare the next model request Sensor tool wrappers1. Pre: budget / hard-block / state_before2. Post: progress / budget settlement3. Detect repetition and stalled work4. Pending signals flow into next model_call Source-specific extraction1. Web, local, or skill reading windows2. Check source, scope, and duplication3. Extract facts with source provenance4. Batch extraction; finalize on exit5. EvidenceGraph + Coverage backfill SOCM State ChangeEvidenceNode / Coverage fill / alias
Case

Case Study

Follow one question on its journey — from the moment it arrives to the finished report.md. Five milestones, one of them a loop.

1

Question In entry

A question arrives. SearchOS spins up a fresh workspace, starts with a blank slate of state, and wakes the Orchestrator — the single mind that will run the whole search.

2

Plan & Map decompose

The Orchestrator turns the question into a plan. When things are vague it sends a scout ahead first; then it draws a coverage map — a table of exactly which facts still need to be found.

3

Search Loop repeats

The heart of it. Several search agents fan out in parallel, read pages, and evidence is captured automatically into the graph and the map. The Orchestrator checks progress and keeps dispatching — round after round — until coverage is enough.

4

Write Up synthesize

Once the map is full enough, the writer agent reads the collected evidence and drafts the answer section by section — every claim tied back to its sources, conflicts flagged rather than hidden.

5

Report Out deliver

The harness assembles the final document and always guarantees a report.md — falling back through simpler synthesis modes if needed, so a session never ends empty-handed.

Along the way, dead-end and low-yield searches are quietly distilled into a failure memory — so the system, and future sessions, stop repeating the same mistakes.
report.md real run 1 / 4
Q

user · widesearch benchmark“My son is starting postgrad applications — find the top five universities in each of the five broad QS subjects (2025), with their QS & THE world ranks, official homepage, application deadline and fee. Put it all in one table.”

200/200cells filled
342evidence
13sources
18rounds
~6minruntime
SubjectUniversityQS·SubjQSTHEDeadlineFee
Arts & HumanitiesHarvard University111Jan 1$90
Arts & HumanitiesStanford University466Jan 5$100
Arts & HumanitiesUniversity of California, Berkeley (UCB)5128Nov 30$80
Arts & HumanitiesUniversity of Cambridge355Oct 15$80
Arts & HumanitiesUniversity of Oxford231Oct 15◇ $36-$60（部分）
Engineering & TechnologyETH Zurich - Swiss Federal Institute of Technology5711Mar 31$187
Engineering & TechnologyMassachusetts Institute of Technology (MIT)112Jan 5$75
Engineering & TechnologyStanford University366Jan 5$100
Engineering & TechnologyUniversity of California, Berkeley (UCB)6128Nov 30$80
Engineering & TechnologyUniversity of Cambridge455Oct 15$80
Life Sciences & MedicineHarvard University111Jan 1$90
Life Sciences & MedicineJohns Hopkins University33216Jan 2$70
Life Sciences & MedicineMassachusetts Institute of Technology (MIT)512Jan 5$75
Life Sciences & MedicineStanford University466Jan 5$100
Life Sciences & MedicineUniversity of Oxford231Oct 15◇ $36-$60（部分）
Natural SciencesHarvard University111Jan 1$90
Natural SciencesMassachusetts Institute of Technology (MIT)212Jan 5$75
Natural SciencesStanford University566Jan 5$100
Natural SciencesUniversity of Cambridge455Oct 15$80
Natural SciencesUniversity of Oxford331Oct 15◇ $36-$60（部分）
Social Sciences & ManagementHarvard University111Jan 1$80
Social Sciences & ManagementMassachusetts Institute of Technology (MIT)412Jan 5$75
Social Sciences & ManagementStanford University366Jan 5$90
Social Sciences & ManagementUniversity of Cambridge555Oct 15$80
Social Sciences & ManagementUniversity of Oxford231Oct 15◇ $36-$60（部分）
25 rows
Every cell is cited — sources include
topuniversities.com timeshighereducation.com harvard.edu mitadmissions.org cam.ac.uk ethz.ch jhu.edu +6 more
Q

user · widesearch benchmark“I’m a huge fan of Lana Del Rey — list every song on her official studio albums (from Born to Die through Tunnel Under Ocean Blvd), standard editions, with release years. One table.”

2schemas
8→109albums→songs
286evidence
150llm calls
~6minruntime
albums schema ① 8 rows
albumrelease_year
Blue Banisters2021
Born to Die2012
Chemtrails over the Country Club2021
Did You Know That There's a Tunnel Under Ocean Blvd2023
Honeymoon2015
Lust for Life2017
Norman Fucking Rockwell!2019
Ultraviolence2014
one album → many songs, joined on album
songs schema ② 109 rows
albumrelease_yearsong
Blue Banisters2021Arcadia
Blue Banisters2021Beautiful
Blue Banisters2021Black Bathing Suit
Blue Banisters2021Blue Banisters
Blue Banisters2021Cherry Blossom
Blue Banisters2021Dealer
Blue Banisters2021If You Lie Down with Me
Blue Banisters2021Interlude – The Trio
Blue Banisters2021Living Legend
Blue Banisters2021Nectar of the Gods
Blue Banisters2021Sweet Carolina
Blue Banisters2021Text Book
Blue Banisters2021Thunder
Blue Banisters2021Violets for Roses
Blue Banisters2021Wildflower Wildfire
Born to Die2012Blue Jeans
Born to Die2012Born to Die
Born to Die2012Carmen
Born to Die2012Dark Paradise
Born to Die2012Diet Mountain Dew
Born to Die2012Million Dollar Man
Born to Die2012National Anthem
Born to Die2012Off to the Races
Born to Die2012Radio
Born to Die2012Summertime Sadness
Born to Die2012This Is What Makes Us Girls
Born to Die2012Video Games
Chemtrails over the Country Club2021Breaking Up Slowly (featuring Nikki Lane)
Chemtrails over the Country Club2021Chemtrails over the Country Club
Chemtrails over the Country Club2021Dance Till We Die
Chemtrails over the Country Club2021Dark but Just a Game
Chemtrails over the Country Club2021For Free (featuring Zella Day and Weyes Blood)
Chemtrails over the Country Club2021Let Me Love You like a Woman
Chemtrails over the Country Club2021Not All Who Wander Are Lost
Chemtrails over the Country Club2021Tulsa Jesus Freak
Chemtrails over the Country Club2021White Dress
Chemtrails over the Country Club2021Wild at Heart
Chemtrails over the Country Club2021Yosemite
Did You Know That There's a Tunnel Under Ocean Blvd2023A&W
Did You Know That There's a Tunnel Under Ocean Blvd2023Candy Necklace (featuring Jon Batiste)
Did You Know That There's a Tunnel Under Ocean Blvd2023Did You Know That There's a Tunnel Under Ocean Blvd
Did You Know That There's a Tunnel Under Ocean Blvd2023Fingertips
Did You Know That There's a Tunnel Under Ocean Blvd2023Fishtail
Did You Know That There's a Tunnel Under Ocean Blvd2023Grandfather Please Stand on the Shoulders of My Father While He's Deep-Sea Fishing
Did You Know That There's a Tunnel Under Ocean Blvd2023Jon Batiste Interlude
Did You Know That There's a Tunnel Under Ocean Blvd2023Judah Smith Interlude
Did You Know That There's a Tunnel Under Ocean Blvd2023Kintsugi
Did You Know That There's a Tunnel Under Ocean Blvd2023Let the Light In (featuring Father John Misty)
Did You Know That There's a Tunnel Under Ocean Blvd2023Margaret (featuring Bleachers)
Did You Know That There's a Tunnel Under Ocean Blvd2023Paris, Texas (featuring SYML)
Did You Know That There's a Tunnel Under Ocean Blvd2023Peppers (featuring Tommy Genesis)
Did You Know That There's a Tunnel Under Ocean Blvd2023Sweet
Did You Know That There's a Tunnel Under Ocean Blvd2023Taco Truck x VB
Did You Know That There's a Tunnel Under Ocean Blvd2023The Grants
Honeymoon201524
Honeymoon2015Art Deco
Honeymoon2015Burnt Norton (Interlude)
Honeymoon2015Don't Let Me Be Misunderstood
Honeymoon2015Freak
Honeymoon2015God Knows I Tried
Honeymoon2015High by the Beach
Honeymoon2015Honeymoon
Honeymoon2015Music to Watch Boys To
Honeymoon2015Religion
Honeymoon2015Salvatore
Honeymoon2015Swan Song
Honeymoon2015Terrence Loves You
Honeymoon2015The Blackest Day
Lust for Life201713 Beaches
Lust for Life2017Beautiful People Beautiful Problems (featuring Stevie Nicks)
Lust for Life2017Change
Lust for Life2017Cherry
Lust for Life2017Coachella – Woodstock in My Mind
Lust for Life2017Get Free
Lust for Life2017God Bless America – and All the Beautiful Women in It
Lust for Life2017Groupie Love (featuring ASAP Rocky)
Lust for Life2017Heroin
Lust for Life2017In My Feelings
Lust for Life2017Love
Lust for Life2017Lust for Life (featuring the Weeknd)
Lust for Life2017Summer Bummer (featuring ASAP Rocky and Playboi Carti)
Lust for Life2017Tomorrow Never Came (featuring Sean Ono Lennon)
Lust for Life2017When the World Was at War We Kept Dancing
Lust for Life2017White Mustang
Norman Fucking Rockwell!2019Bartender
Norman Fucking Rockwell!2019California
Norman Fucking Rockwell!2019Cinnamon Girl
Norman Fucking Rockwell!2019Doin' Time
Norman Fucking Rockwell!2019Fuck It I Love You
Norman Fucking Rockwell!2019Happiness Is a Butterfly
Norman Fucking Rockwell!2019Hope Is a Dangerous Thing for a Woman Like Me to Have – but I Have It
Norman Fucking Rockwell!2019How to Disappear
Norman Fucking Rockwell!2019Love Song
Norman Fucking Rockwell!2019Mariners Apartment Complex
Norman Fucking Rockwell!2019Norman Fucking Rockwell
Norman Fucking Rockwell!2019The Greatest
Norman Fucking Rockwell!2019The Next Best American Record
Norman Fucking Rockwell!2019Venice Bitch
Ultraviolence2014Brooklyn Baby
Ultraviolence2014Cruel World
Ultraviolence2014Fucked My Way Up to the Top
Ultraviolence2014Money Power Glory
Ultraviolence2014Old Money
Ultraviolence2014Pretty When You Cry
Ultraviolence2014Sad Girl
Ultraviolence2014Shades of Cool
Ultraviolence2014The Other Woman
Ultraviolence2014Ultraviolence
Ultraviolence2014West Coast
109 songs
Every cell is cited — sources include
Lana_Del_Rey_discography Born_to_Die Ultraviolence Honeymoon lanadelrey.fandom.com +4 more
Q

user · widesearch benchmark（中文）“在写美国政治论文，帮我罗列从美国建国到 2025 年大选年份获胜的总统（中文名），表头还要含大选年份、党派、民众得票数、选举人团票数。整理成一整张 Markdown 表。”

60/60cells filled
588evidence
8sources
299llm calls
~9minruntime
年份姓名党派民众票数团票数
1789乔治·华盛顿无党籍—69
1792乔治·华盛顿无党派11176132
1796约翰·亚当斯联邦党3517471
1800托马斯·杰斐逊民主共和党4551173
1804托马斯·杰斐逊民主共和党106786162
1808詹姆斯·麦迪逊民主共和党124732122
1812詹姆斯·麦迪逊民主共和党140431128
1816詹姆斯·门罗民主共和党76762183
1820詹姆斯·门罗民主共和党85443231
1824约翰·昆西·亚当斯民主共和党12244084
1828安德鲁·杰克逊民主党638348178
1832安德鲁·杰克逊民主党701780219
1836马丁·范布伦民主党764176170
1840威廉·亨利·哈里森辉格党1275390234
1844詹姆斯·诺克斯·波尔克民主党1339494170
1848扎卡里·泰勒辉格党1361396163
1852富兰克林·皮尔斯民主党1604167254
1856詹姆斯·布坎南民主党1836072174
1860亚伯拉罕·林肯共和党1855276180
1864亚伯拉罕·林肯国家联盟党2218388212
1868尤利西斯·S·格兰特共和党3013421214
1872尤利西斯·S·格兰特共和党3595319286
1876拉瑟福德·伯查德·海斯共和党4034311185
1880詹姆斯·艾布拉姆·加菲尔德共和党4453336214
1884格罗弗·克利夫兰民主党4909544219
1888本杰明·哈里森共和党5440338233
1892格罗弗·克利夫兰民主党5556917277
1896威廉·麦金莱共和党7112138271
1900威廉·麦金莱共和党7219193292
1904西奥多·罗斯福共和党7625599336
1908威廉·霍华德·塔夫脱共和党7676598321
1912伍德罗·威尔逊民主党6294327435
1916伍德罗·威尔逊民主党9126063277
1920沃伦·甘梅利尔·哈定共和党16151916404
1924卡尔文·柯立芝共和党15724350382
1928赫伯特·胡佛共和党21402344444
1932富兰克林·罗斯福民主党22392711472
1936富兰克林·罗斯福民主党27751235523
1940富兰克林·德拉诺·罗斯福民主党27313945449
1944富兰克林·罗斯福民主党25612608432
1948哈里·S·杜鲁门民主党24179347303
1952德怀特·大卫·艾森豪威尔共和党34075529442
1956德怀特·大卫·艾森豪威尔共和党35579180457
1960约翰·肯尼迪民主党34220984303
1964林登·约翰逊民主党43127041486
1968理查德·尼克松共和党31783783301
1972理查德·尼克松共和党47169911520
1976吉米·卡特民主党40830763297
1980罗纳德·里根共和党43904153489
1984罗纳德·里根共和党54455075525
1988乔治·H·W·布什共和党48886097426
1992比尔·克林顿民主党44908254370
1996比尔·克林顿民主党47402357379
2000乔治·沃克·布什共和党50456002271
2004乔治·沃克·布什共和党62040610286
2008贝拉克·侯赛因·奥巴马民主党69498516365
2012巴拉克·歐巴馬民主党65915795332
2016唐纳·川普共和党62984828304
2020乔·拜登民主党81283501306
2024唐纳·川普共和党77304184312
60 届大选
每个单元格均有引用 · 来源包括
zh.wikipedia.org presidency.ucsb.edu 270towin.com archives.gov britannica.com +3 more
Q

user · widesearch benchmark（中文）“在做女性力量成长轨迹研究，想看看获得诺贝尔文学奖的女性都有哪些，以及她们最早出版的三部作品（姓名与作品名保留原文）。整理成一张表。”

2schemas
18→52作家→作品
264evidence
383llm calls
~10minruntime
诺贝尔文学奖女性获奖者 schema ① 18 位
姓名获奖年份
Alice Munro2013
Annie Ernaux2022
Doris Lessing2007
Elfriede Jelinek2004
Gabriela Mistral1945
Grazia Deledda1926
Han Kang2024
Herta Müller2009
Louise Glück2020
Nadine Gordimer1991
Nelly Sachs1966
Olga Tokarczuk2018
Pearl S. Buck1938
Selma Lagerlöf1909
Sigrid Undset1928
Svetlana Alexievich2015
Toni Morrison1993
Wisława Szymborska1996
一位作家 → 三部早期作品，按 姓名 关联
作家早期作品 schema ② 52 部
姓名出版作品名称首次出版年份
Alice MunroDance of the Happy Shades1968
Alice MunroThe View from Castle Rock1977
Alice MunroWho Do You Think You Are?1978
Annie ErnauxLes Armoires vides1974
Annie ErnauxCe qu'ils disent ou rien1977
Annie ErnauxLa Femme gelée1981
Doris LessingThe Grass Is Singing1950
Doris LessingMartha Quest1951
Doris LessingA Proper Marriage1954
Elfriede JelinekLisas Schatten1967
Elfriede JelinekIm Moor1973
Elfriede Jelinekende / gedichte von 1966–19682000
Gabriela MistralDesolación1922
Gabriela MistralTernura: canciones de niños1922
Gabriela MistralTernura1924
Grazia DeleddaNell'azzurro1890
Grazia DeleddaFior di Sardegna1891
Grazia DeleddaPaesaggi sardi1896
Han Kang——
Herta MüllerNiederungen1982
Herta MüllerDrückender Tango1984
Herta MüllerDer Mensch ist ein großer Fasan auf der Welt1986
Louise GlückFirstborn1968
Louise GlückThe House on Marshland1975
Louise GlückDescending Figure1980
Nadine GordimerFace to Face1949
Nadine GordimerThe Soft Voice of the Serpent and Other Stories1952
Nadine GordimerThe Lying Days1953
Nelly SachsLegenden und Erzählungen1921
Nelly SachsIn den Wohnungen des Todes1946
Nelly SachsSternverdunkelung1949
Olga TokarczukPodróż ludzi Księgi1993
Olga TokarczukE.E.1995
Olga TokarczukPrawiek i inne czasy1996
Pearl S. BuckEast Wind: West Wind1930
Pearl S. BuckThe Good Earth1931
Pearl S. BuckSons1932
Selma LagerlöfGösta Berlings saga1891
Selma LagerlöfOsynliga länkar1894
Selma LagerlöfAntikrists mirakler1897
Sigrid UndsetFru Marta Oulie1907
Sigrid UndsetDen lykkelige alder1908
Sigrid UndsetFortællingen om Viga-Ljot og Vigdis1909
Svetlana AlexievichУ войны не женское лицо1984
Svetlana AlexievichПоследние свидетели1985
Svetlana AlexievichЦинковые мальчики1989
Toni MorrisonThe Bluest Eye1970
Toni MorrisonSula1973
Toni MorrisonSong of Solomon1977
Wisława SzymborskaDlatego żyjemy1952
Wisława SzymborskaPytania zadawane sobie1954
Wisława SzymborskaWołanie do Yeti1957
52 部
每个单元格均有引用 · 来源包括
nobelprize.org en.wikipedia.org isfdb.org ko.wikipedia.org fr.wikipedia.org +4 more
single-schema · 25×8 table
Results

Benchmark Results

SearchOS-V1 results from the paper, covering Web research benchmarks.

80.3
WideSearch · Item F1
56.5
WideSearch · Row F1
76.9
GISA · Table F1
76.5
GISA · Set F1
Benchmark Metric Single Agent Multi‑Agent System
ReAct Plan‑and‑Solve Table‑as‑Search A‑MapReduce Web2BigTable Ours
WideSearch Item · Precision 82.983.882.483.178.383.9
Item · Recall 70.272.973.574.273.479.7
Item · F1 72.975.275.476.073.880.3
Row · Precision 58.058.757.156.957.559.0
Row · Recall 48.850.251.649.854.055.8
Row · F1 50.952.252.751.454.556.5
GISA Table · Item · F1 74.871.273.472.568.176.9
Table · Row · F1 58.150.754.152.145.359.7
Set · F1 61.663.160.962.556.776.5
List · F1 67.153.854.257.465.568.1
Item · EM 0.016.716.733.350.050.0
max@3 — best of three runs per case, ×100. Each row’s best is highlighted; the Ours column is tinted. Item scores cells independently; Row requires a fully correct row.
Resources

Resources & Citation

Read the Paper View on GitHub
BibTeX 
@article{zhang2026searchos,
  title={SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration},
  author={Zhang, Yuyao and Gao, Junjie and Wu, Zhengxian and Fan, Jiaming and Zhang, Jin and Ma, Shihan and Yao, Yao and Qi, Weiran and Jin, Chuyan and Ma, Guiyu and others},
  journal={arXiv preprint arXiv:2607.15257},
  year={2026}
}
Credits

Authors & Acknowledgements

Yuyao Zhang1,2,*,‡ Junjie Gao2,* Zhengxian Wu2 Jiaming Fan2 Jin Zhang2 Shihan Ma2 Yao Yao2 Weiran Qi2 Guiyu Ma2 Xingzhong Xu2 Kai Yang2 Ji-Rong Wen1 Zhicheng Dou1,†

1 Renmin University of China 2 Ant Group

SearchOS was developed through the joint efforts of the authors and contributors listed above. We thank Ant Insurance for its strong support throughout the project.