suvyo.
Work with us
DOCUMENTATION

Understand how it works.

The contracts behind the catalog, evidence and contribution workflows. Clear enough to inspect, small enough to operate.

Getting started

Use Discover to search stored resource records. Open a detail page to inspect source attribution, license metadata, revision and synchronization time. Select models in Compare to view specifications side by side.

Sign in with ChatGPT to submit a resource or evaluation request. Your account shows review status and lets you withdraw contributions. Administrative roles require an explicit server-side allowlist.

Discovery & ranking

Search runs against the persisted catalog, so an upstream outage does not require a live provider request for every visit. Exact and partial name matches, task matches and description terms inform relevance. Popularity is not treated as correctness.

Filters include type, task, language, license, source, modality, Indian-language relevance, local deployment indication and open-weight indication. Filter values and pagination live in the URL. Missing source fields are displayed as unknown.

Provenance & licenses

Provider identity and source payloads are stored alongside Suvyo-specific metadata. The initial sources are the Hugging Face Hub API, GitHub REST API and arXiv metadata feeds. Source URLs, revisions, update times and fetch times remain visible.

A source-declared license is a publisher statement. It is not a legal opinion or a guarantee that all files in a repository share that license. Public accessibility does not imply open-source permission. Gated files are not bypassed or mirrored.

AIKosh is linked as an ecosystem resource; an authorized metadata integration has not been verified. Read the attribution policy.

Suvyo Proof

The first suite is a small, versioned English/Hindi tool-routing and structured-output benchmark. Each input has an expected route and an output schema. The runner validates JSON syntax, schema shape and task correctness separately, preserving raw responses and timing.

Execution records carry a model/checkpoint ID, immutable model revision, dataset hash, code hash, prompt/runtime settings, start/end times and artifact hash. Test-fixture runs are explicitly synthetic and rejected by publication controls.

Run the documented CLI in the repository against a permitted local OpenAI-compatible inference endpoint. An endpoint being available does not establish which model revision it serves; the operator must verify it and supply complete runtime evidence.

node evaluation/runner.mjs --help

A reviewer must inspect a completed artifact before publication. Different suite versions, datasets or runtime settings must not be collapsed into one leaderboard. A routing benchmark does not establish broad language quality or production safety.

Contribution workflow

  1. Sign in and submit a title, original HTTPS URL, type and useful description.
  2. The server validates size, link safety, duplication and ownership.
  3. An administrator reviews evidence and records an approval or rejection.
  4. Approved publication retains source attribution. Submitted claims do not receive evaluation badges.
  5. Use your account to inspect feedback or withdraw your submission.

Only submit content you are entitled to share. Do not include secrets or private datasets. Follow upstream maintainers’ contribution rules. Community guidelines explain the review expectations.

Read-only catalog API

GET /api/v1/resources?q=translation&type=model&page=1&limit=12
GET /api/v1/resources/{slug}
GET /api/v1/compare?ids={slug},{slug}
GET /api/v1/stats
GET /api/v1/content?type=articles
GET /api/v1/proof

Resource responses contain items, total, page, limit, pages and facets. Records retain stable IDs, source identity and provenance. Detail routes return an error for unknown resources.

Errors use {"error":{"code":"…","message":"…"}}. Expect 400 for invalid requests, 401 for missing identity, 403 for denied access, 404 for unknown records, 429 for rate limiting and 503 when a dependent service is unavailable. Read responses have short caches. Authentication and a private Site’s access boundary still apply.

Use small pages, cache results, honor 429 responses and retain attribution. Private user, inquiry and moderation records are never included in this public API.

Integration guidance

Use original providers to obtain weights or datasets under their terms. Suvyo supplies discovery metadata and evidence, not an inference endpoint. Keep tool execution behind application-level validation and permissions. Evaluate the exact model revision, language and task you intend to deploy.

Provider ingestion is bounded, timeout-controlled and restricted to known API hosts. New discoveries enter review. Scheduled synchronization refreshes existing source records; it does not certify model quality.

Roadmap & changelog

Phase 1, 23 October 2026: source-linked discovery, model comparison, authored research, proposed research tracks, opportunity briefs, contribution review, inquiry handling and the first executable Proof suite.

Next priorities are broader permitted model executions, reviewed Indic examples, dependable source freshness and operational validation. Suvyo Route, Data, Code and Bhasha remain proposed tracks.

The platform’s policies are founder-review drafts. No claim of formal legal compliance or general model safety is made.