feat(tinyfish): make search provider permissive, attribute errors

[ChenluJi]

Summary

Follow-up to #30634. Reshapes the TinyFish search provider so LiteLLM mirrors the TinyFish Search API surface instead of maintaining a parallel cherry-pick.

Request side

1. Drop the request TypedDict. Was a runtime no-op and misleadingly listed include_thumbnail. Removed; readers should refer to TinyFish's own docs.
2. Stop sending max_results on the wire. TinyFish doesn't honor it server-side; clamp to [1, 10] (TinyFish's natural SERP cap) and thread the caller value through self for client-side response truncation.
3. Guard max_results="abc". Previously raised a bare ValueError; now logs a warning and treats as unset.
4. Auto-JSON-encode dict params; lowercase bool serialization. Callers naturally pass fetch={"format": "html"} (dict) and include_thumbnail=True (bool); serialize both before urlencode so the existing strict adapter and ux-labs' literal-"true"/"false" validator accept them.

Response side

5. Drop both Pydantic response models. Parse directly into LiteLLM's SearchResponse. Per-result extras (position, site_name, fetch, fetch_error, etc.) flow through automatically via SearchResult's extra="allow".
6. Default missing per-result title/url/snippet to "". A degraded result flows through instead of failing the whole call.
7. Read TinyFish's parameter_warnings field when present and re-fire each entry as a verbose_logger.warning. Defensive on every malformed shape (non-list field, non-dict entry, missing required keys, non-string types). Pre-wired for a future TinyFish-side rollout; no-op today.

Error handling

8. Inline attributed error wrapping (_wrap_error). Three call sites in transform_search_response (non-2xx, JSONDecodeError, ValidationError) wrap errors with "TinyFish Search: <msg>. See https://docs.tinyfish.ai/search-api for details." Network failures and "no response after retries" flow through LiteLLM core's _handle_error → base get_error_class and inherit any future upstream improvements; their attribution is implicit via the hostname in the bare error message.
9. Dispatch non-2xx responses through _wrap_error at the top of transform_search_response. AsyncHTTPHandler.get doesn't call raise_for_status, so 4xx/5xx bodies were previously parsed as if successful — a pre-existing bug that silently returned SearchResponse(results=[]) on auth failures, rate limits, etc.
10. Wrap json.JSONDecodeError on 200 bodies through the same path.
11. Wrap pydantic.ValidationError for envelope-shape mismatches so callers see "TinyFish Search: Response shape does not match LiteLLM's SearchResponse schema: ..." instead of an opaque pydantic traceback.

Behavior changes worth flagging

• 4xx/5xx responses now raise an attributed BaseLLMException instead of silently returning SearchResponse(results=[]). Pre-existing bug fix.

Test plan

• 58 unit + integration tests pass (tests/test_litellm/llms/tinyfish/, tests/search_tests/test_tinyfish_search.py)
• Full LLM-provider suite passes (sibling providers unaffected — 7170+ passing; pre-existing bedrock/custom_httpx flakes unrelated)
• Ruff clean
• No new basedpyright errors beyond the 4 pre-existing in this file (new errors all in low-priority budget categories with ample slack)
• Live-tested against the production TinyFish endpoint: 47 cases covering happy paths, all unified params, all TinyFish-side params (page, recency_minutes, after_date, before_date, fetch, include_thumbnail, language, location), boundary values, cross-field validation, length limits, malformed inputs, auth failures, network failures, and combinations. All wrap or pass as expected.

[CLAassistant]

All committers have signed the CLA.

[ChenluJi]

@greptileai review

[greptile-apps]

Greptile Summary

This PR reshapes the TinyFish search provider to mirror TinyFish's actual API surface: removes a now-redundant TypedDict and two Pydantic response models, stops sending max_results on the wire (TinyFish ignores it server-side), and adds inline attributed error wrapping so 4xx/5xx responses raise a BaseLLMException instead of silently returning empty results.

• Request side: transform_search_request now threads max_results through self._caller_max_results for client-side slicing (clamped to TinyFish's 10-result SERP ceiling), auto-JSON-encodes dict params (e.g. fetch), and lowercases Python bool params to satisfy ux-labs' validator.
• Response side: Parses directly into SearchResponse (leveraging extra="allow" for per-result extras), defaults missing title/url/snippet to "" in-place before validation, and reads parameter_warnings defensively with per-entry validation.
• Error handling: _wrap_error unwraps ux-labs' {"error": {"message": ...}} envelope and raises BaseLLMException with a consistent "TinyFish Search: … See docs …" prefix for non-2xx, JSONDecodeError, and ValidationError paths.

Confidence Score: 5/5

Safe to merge — the changes fix a real bug (silent 4xx/5xx swallowing), align the wire format with TinyFish's actual behavior, and are covered by a thorough test suite with no weakened assertions.

The transformation logic is correct: state threading via _caller_max_results is safe given per-call instantiation, _default_missing_result_fields correctly defaults before pydantic validation, _wrap_error handles all known ux-labs envelope shapes defensively, and _emit_parameter_warnings is fully guarded against malformed shapes. The only finding is a stale test helper in the integration test file that doesn't affect any current tests.

tests/search_tests/test_tinyfish_search.py — the _make_mock_response helper was not updated to set mock.text and mock.headers, which the production code now accesses in error paths. Harmless today (no error-path tests in that file), but worth aligning.

Important Files Changed

Filename	Overview
litellm/llms/tinyfish/search/transformation.py	Comprehensive refactor: removes custom TypedDicts/Pydantic models, adds proper non-2xx error attribution via _wrap_error, threads max_results through instance state for client-side truncation, auto-encodes dict/bool params, and adds defensive parameter_warnings forwarding
tests/test_litellm/llms/tinyfish/test_tinyfish_search.py	Well-updated unit test suite: _make_mock_response helper now supports text, headers, and json_data=None; new error-handling tests cover 4xx, 429, 5xx, non-JSON, and schema-mismatch paths; new request-side tests cover dict/bool serialization, max_results wire exclusion, and per-param pass-through
tests/search_tests/test_tinyfish_search.py	Integration-style test file: adds test_fetch_param_round_trip for end-to-end dict param serialization, but _make_mock_response is not updated to set mock.text or mock.headers, which are both now accessed by the production code in error paths

_{Reviews (5): Last reviewed commit: "fix(tinyfish): reduce transform_search_r..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[ChenluJi]

@greptileai review

[ChenluJi]

@greptileai review

[ChenluJi]

@greptileai review