[Integration] infiniflow/ragflow — You.com web search provider

[brainsparker]

Target Repository URL

https://github.com/infiniflow/ragflow

You.com API(s)

search

Estimated Complexity

large

Integration Description

RAGFlow (88.8K★, InfiniFlow) wires web search into both chat retrieval and the agent tool set. The web_search_provider field selects the provider per chat, with BYOK credentials entered in the chat settings UI.

This integration adds You.com as a selectable provider across all four layers: the Go chat/deep-research runtime, the Python retrieval runtime, the chat settings UI, and the HTTP API docs.

Target Repo Contributing Guidelines

No CONTRIBUTING.md in the repo. Issue templates exist; PRs are reviewed by CodeRabbit.

Planned Approach

Mirror the two providers already wired behind the abstraction:

• internal/service/web_search_provider.go — webSearchProviderYouCom, endpoint selection, and retrieveYouComWebSearch returning RAGFlow's chunk/doc_agg shape.
• rag/utils/youcom_conn.py — YouCom connector; registered in rag/utils/web_search_conn.py.
• Web: constants/chat.ts, interfaces/database/chat.ts, components/web-search-form-field.tsx, pages/next-chats/chat/{web-search-api-key.ts,use-show-internet.ts,app-settings/use-chat-setting-schema.tsx}, locales/{en,zh}.ts.
• docs/references/http_api_reference.md.
• Tests in Go, Python, and the web suite.

Keyless: no youcom_api_key → api.you.com/v1/agents/search; key set → api.you.com/v1/search with X-API-Key.

Status Checklist

• Claimed this integration (self-assign)
• Filed issue or opened discussion in target repo
• Opened PR in target repo
• You.com API calls send the standard integration User-Agent (youdotcom-integration/<owner>-<repo>)
• You.com links in contributed docs carry the standard UTM parameters
• PR merged or rejected

Links

Target repo issue: none — issue-first dropped on evidence (see notes); confirmed with Brian
Target repo PR: infiniflow/ragflow#18478

Additional Notes

Scope correction — Tavily is no longer the only wired provider. The brief described a single concrete implementation behind an unused abstraction, and recommended filing an issue first to confirm they wanted a second provider. Neither holds:

• Querit already landed as a second provider — PR #17813 "feat(chat): add Querit web search provider", merged 2026-08-04, authored by a vendor (ethan7zhanghx), with no preceding issue. It also did the abstraction refactor. Querit followed up with an agent search tool (#17548) and a Contents tool (#18156), both merged.
• A third-provider PR is already open — #18475 "feat(chat): add Serply web search provider", opened 2026-08-19, +664 across 18 files. That PR is the canonical template for a third provider, and it is a live competitor for the same slot.

Issue-first was therefore dropped on the strength of that evidence. Confirmed with Brian before opening the PR.

Our differentiator is keyless. RAGFlow's model is BYOK-per-chat: has_web_search_provider (Python) and resolveWebSearchProvider (Go) both refuse a provider with an empty key, and the UI hides the Internet toggle until a key is entered. You.com is registered as the only keyless-capable provider, so selecting it enables web search with nothing pasted. That required a small, per-provider carve-out in each of those three gates rather than a change to shared behaviour — Tavily, Querit, and Serply are unaffected.

No official You.com logo asset exists at a usable size (Querit ships a 499×472 PNG; you.com serves only a 16×16 .ico). The provider option renders without a logo, with the logo field made optional and a comment marking the follow-up.

---

User-Agent shipped: RAGFlow youdotcom-integration/infiniflow-ragflow (case 1), sent from both the Go and Python runtimes, on both auth paths, You.com hosts only. Pinned by Go and Python tests.

UTM: 2026-08-oss-integrations / utm_source=infiniflow-ragflow / utm_content=app on the "How to get it?" link in the chat-settings provider form.

[brainsparker]

PR opened: infiniflow/ragflow#18478 — feat(chat): add You.com web search provider (+557/−19, 17 files).

Ratchet: Not Started → PR Submitted.

Covers all four layers behind the web_search_provider abstraction: the Go chat + deep-research runtime, the Python retrieval runtime, the chat settings UI (en + zh), and the HTTP API reference. 7 Go tests, 12 Python, 4 frontend.

Keyless is the differentiator and it required real work. RAGFlow is BYOK-per-chat: resolveWebSearchProvider (Go), has_web_search_provider (Python), and the UI's Internet toggle all refuse a provider with an empty key. Each got a per-provider carve-out — a keyOptional flag, a KEYLESS_WEB_SEARCH_PROVIDERS set, and a hasWebSearchProvider helper — so You.com is selectable with nothing pasted while Tavily, Querit, and Serply resolve exactly as before. There are tests asserting the no-change. The PR offers to delete all three carve-outs if the maintainer wants uniformity.

Verification, with one honest gap. Python: 20 passed (9 new connector + 3 new dispatch, existing Querit tests unaffected); ruff clean. Frontend: 12 passed (4 new); oxlint clean; tsc --noEmit at 218 errors both with and without the change (pre-existing, verified by stashing). Go: gofmt clean and all 7 new tests pass — but go build ./internal/service/ cannot run in this environment, failing identically on a clean main checkout (a cgo dep missing its header, and syscall.SysProcAttr.Pdeathsig being Linux-only on macOS). I compiled the new code, its tests, and resolveWebSearchProvider in an isolated module instead — go vet clean, go test green, plus a live run against the real API returning 6 correctly shaped chunks on both endpoints. CI will be the real Go gate. This is stated plainly in the PR body.

Competitive: Serply's #18475 is open against the same dispatch, filed the same day. The PR offers to rebase if it merges first — they touch the same switch statements but no shared lines beyond that.

Follow-up: no usable You.com logo asset exists (Querit ships 499×472; you.com serves only a 16×16 .ico), so providerOptions.logo was made optional and the You.com row renders its name alone with a comment marking the spot. Same asset gap as Scira — worth sourcing one properly.

[brainsparker]

Rebased and review-addressed. PR #18478 is MERGEABLE again.

Serply merged first. Competitor PR infiniflow/ragflow#18475 landed at 10:59Z and conflicted our branch across 7 files. Rebuilt on top of the new main rather than untangling ~20 interleaved hunks, so You.com now sits alongside Serply in every registry instead of racing it. Serply's code untouched. Diff is +955/−18 across 17 files.

CodeRabbit's three findings, all taken:

1. youComContent returned a whitespace-only description unchanged, which could produce a chunk with no usable content → now trimmed, plus TestYouComContentDiscardsBlankSnippetsAndDescriptions (3 cases) and TestRetrieveYouComWebSearchSkipsBlankContent.
2. Raw query and page text in info logs (flagged Major/privacy) → retrieve_chunks now logs counts only. Noted in-thread that querit_conn.py and tavily_conn.py share the pattern — left them as out of scope and offered it as a follow-up sweep rather than diverging silently.
3. API-reference wording implied omitting web_search_provider enables Tavily unconditionally → clarified to match both resolvers.

One test added unprompted, in Go and Python: the keyless carve-out must not relax any other provider. tavily, querit, serply all asserted to still resolve to nothing without a key. That was the only real risk in the diff and it's now pinned.

Re-validated post-rebase: Python 26 passed (Serply's suite unaffected), ruff clean. Frontend 14 passed (incl. Serply's cases), oxlint clean, tsc 219 errors with and without the change. Go gofmt clean, 10 tests pass, live run returns 6 correctly shaped chunks on both endpoints — same standing caveat that the Go package won't build on macOS, so CI is the gate.

---

Separately, and more important than the PR: InfiniFlow co-founder Jin Hai asked us to email haijin.chn@gmail.com — he'll notify us when they cut a release so we can verify You.com works, and wants a fast fix loop from our side for patch releases.

The ticket recorded "no outreach path exists" for this org. That's no longer true — it's inbound. Escalated to Brian in Linear (DX-760); I replied in the PR thread that the request is passed on and someone will be in touch, so it isn't left hanging. Not emailing a founder on our own initiative.

[brainsparker]

Full CI green, then one more privacy fix.

Everything passed on aa5d468 — including both Go jobs (ragflow_tests_elasticsearch (go) 7m13s, ragflow_tests_infinity (go) 6m30s). That closes the verification gap I'd flagged: the Go package won't build on macOS, so CI was the only real gate on that code, and it's clean. Python jobs green too, CodeRabbit review completed, mergeStateStatus: CLEAN.

Logo shipped. Brian supplied the official SVG. The provider row now renders the You.com brand mark at web/src/assets/svg/youcom.svg. Because every provider has a logo again, I reverted the optional-logo plumbing the placeholder had required — providerOptions is back to upstream's shape and that file's diff is now purely additive, one import plus two entries, exactly matching how Serply was added.

New CodeRabbit finding, valid, fixed in 54347bc. requests builds its HTTPError message from the response URL, and this connector passes the query as a URL parameter — so logging str(error) wrote the user's search query into the error log. Now logs the status code for HTTP failures and the exception type for transport failures; _safe_error_message deleted since it only scrubbed the API key from a message we no longer log.

Two regression tests added, and I verified they go red against the previous handler rather than passing vacuously.

Worth knowing: rag/utils/serply_conn.py has the identical leak — same requests.get(..., params=...) plus _safe_error_message pattern, merged earlier today. querit_conn.py is unaffected because it POSTs. Flagged to the maintainer with an offer to fix it in a separate PR. That's the second time a review finding on our file has turned out to apply to their existing code; surfacing it lands well and costs nothing.

Python suite now 27 passed (Serply's and Querit's own tests unaffected). CI re-running on the new commit.

[brainsparker]

CI green on the log-leak fix (54347bc). All six checks pass — CodeRabbit, preflight, and both Go and both Python test jobs. mergeStateStatus: CLEAN, mergeable: MERGEABLE.

All four CodeRabbit findings are resolved with no new ones raised on the latest commit. Nothing outstanding on our side; the PR is waiting on maintainer review.

[brainsparker]

Maintainer replied yes to all five asks. Two were code; both shipped.

PR	What	Size
#18582	connector logging cleanup (querit + tavily + serply)	+102
#18583	You.com agent tool	+594

#18478 remains green and unchanged.

Keyless is now maintainer-approved — "We're fine with your recommended approach — keyless by default with an optional key to lift rate limits." That was the main review risk on #18478 and it's settled.

Agent tool sizing mattered. Querit's equivalent PR was +3527 across 46 files; Keenable's was +392 across 16. The gap is entirely the frontend form — Querit exposes dynamic site/country/language lists needing bespoke utils and watch hooks. Following Keenable and keeping the parameter surface to query + freshness landed us at +594.

Logging cleanup went one file wider than asked. He named querit and tavily; serply had the sharper version of the same bug (GET with the query in the URL, so failures logged the query itself). Verified precisely against upstream: the key was redacted correctly, the query was not.

---

Flagged separately to Brian: /v1/agents/search degraded for ~15 minutes. 7/20 requests succeeded (connection failures and 503s) while /v1/search stayed 20/20 with the same client and network. Recovered since — 12/12 on recheck.

Recording it here because keyless is the differentiator in all four integrations in this batch, and the failure was asymmetric: keyed unaffected, keyless at 35%. Worth watching whether the free tier is systematically less reliable, since that's the path a maintainer evaluating us hits by default.