Skip to content

Supercharge your AI agents with data from the web and beyond. Building the library for superintelligence. 🔥

feat(api): move query format to gemini-3.5-flash-lite with minimal thinking (#4901)

Gergő Móriczfd9ead29ad04
diff --git a/apps/api/src/__tests__/snips/v2/scrape-query.test.ts b/apps/api/src/__tests__/snips/v2/scrape-query.test.ts
index 6bf0e3d87..3cd9ffb39 100644
--- a/apps/api/src/__tests__/snips/v2/scrape-query.test.ts
+++ b/apps/api/src/__tests__/snips/v2/scrape-query.test.ts
@@ -64,6 +64,28 @@ describe("Query format", () => {
scrapeTimeout,
);
+ concurrentIf(TEST_PRODUCTION || (HAS_AI && ALLOW_TEST_SUITE_WEBSITE))(
+ "answers a query from the page content",
+ async () => {
+ const response = await scrape(
+ {
+ url: `${TEST_SUITE_WEBSITE}/product`,
+ formats: [
+ {
+ type: "query",
+ prompt: "What is the price of the Firecrawl Test Widget?",
+ },
+ ],
+ },
+ identity,
+ );
+
+ expect(response.answer).toContain("49.99");
+ expect(response.warning ?? "").not.toContain("Query generation failed");
+ },
+ scrapeTimeout,
+ );
+
concurrentIf(TEST_PRODUCTION || (HAS_AI && ALLOW_TEST_SUITE_WEBSITE))(
"returns both answer and markdown when formats include markdown and query",
async () => {
@@ -164,6 +186,25 @@ describe("Query format", () => {
scrapeTimeout,
);
+ it(
+ "rejects an unknown query mode",
+ async () => {
+ const response = await scrapeWithFailure(
+ {
+ url: TEST_SUITE_WEBSITE,
+ formats: [
+ { type: "query", prompt: "What is Firecrawl?", mode: "summary" },
+ ],
+ } as any,
+ identity,
+ );
+
+ expect(response.success).toBe(false);
+ expect(response.error).toBeDefined();
+ },
+ scrapeTimeout,
+ );
+
it(
"rejects question over 10000 characters",
async () => {
diff --git a/apps/api/src/lib/extract/usage/model-prices.ts b/apps/api/src/lib/extract/usage/model-prices.ts
index c100af736..501c34291 100644
--- a/apps/api/src/lib/extract/usage/model-prices.ts
+++ b/apps/api/src/lib/extract/usage/model-prices.ts
@@ -19405,6 +19405,21 @@ export const modelPrices = {
supports_tool_choice: true,
supports_web_search: true,
},
+ "gemini-3.5-flash-lite": {
+ max_tokens: 65536,
+ max_input_tokens: 1048576,
+ max_output_tokens: 65536,
+ input_cost_per_token: 3e-7,
+ output_cost_per_token: 2.5e-6,
+ output_cost_per_reasoning_token: 2.5e-6,
+ cache_read_input_token_cost: 3e-8,
+ litellm_provider: "vertex_ai-language-models",
+ mode: "chat",
+ supports_reasoning: true,
+ supports_system_messages: true,
+ source:
+ "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing",
+ },
"vertex/gemini-3.1-flash-lite": {
max_tokens: 65536,
max_input_tokens: 1048576,
diff --git a/apps/api/src/scraper/scrapeURL/transformers/calculateCost.test.ts b/apps/api/src/scraper/scrapeURL/transformers/calculateCost.test.ts
index 13cf63212..5890367c2 100644
--- a/apps/api/src/scraper/scrapeURL/transformers/calculateCost.test.ts
+++ b/apps/api/src/scraper/scrapeURL/transformers/calculateCost.test.ts
@@ -13,6 +13,7 @@ describe("calculateCost", () => {
["openai/o3-mini", 1.1, 4.4],
["gemini-2.5-flash-lite", 0.1, 0.4],
["google/gemini-2.5-flash-lite", 0.1, 0.4],
+ ["gemini-3.5-flash-lite", 0.3, 2.5],
["gpt-5-mini", 0.25, 2],
["fireworks_ai/accounts/fireworks/models/gpt-oss-20b", 0.05, 0.2],
])("prices %s", (model, inputPerM, outputPerM) => {
diff --git a/apps/api/src/scraper/scrapeURL/transformers/query.test.ts b/apps/api/src/scraper/scrapeURL/transformers/query.test.ts
index eb2070b61..9c4bc3d09 100644
--- a/apps/api/src/scraper/scrapeURL/transformers/query.test.ts
+++ b/apps/api/src/scraper/scrapeURL/transformers/query.test.ts
@@ -209,7 +209,10 @@ describe("performQuery freeform", () => {
);
const [gemini, mini] = calls();
- expect(gemini.model.modelId).toBe("gemini-2.5-flash-lite");
+ expect(gemini.model.modelId).toBe("gemini-3.5-flash-lite");
+ expect(gemini.providerOptions).toEqual({
+ google: { thinkingConfig: { thinkingLevel: "minimal" } },
+ });
expect(gemini.prompt).toContain(markdown);
expect(mini.model.modelId).toBe("gpt-4o-mini");
// The page is trimmed to exactly 80% of gpt-4o-mini's window; the rest of
diff --git a/apps/api/src/scraper/scrapeURL/transformers/query.ts b/apps/api/src/scraper/scrapeURL/transformers/query.ts
index 2c211d565..2e579ef88 100644
--- a/apps/api/src/scraper/scrapeURL/transformers/query.ts
+++ b/apps/api/src/scraper/scrapeURL/transformers/query.ts
@@ -1,4 +1,5 @@
import { generateText } from "ai";
+import type { GoogleLanguageModelOptions } from "@ai-sdk/google";
import { encoding_for_model } from "@dqbd/tiktoken";
import { Document, FormatObject } from "../../../controllers/v2/types";
import { Meta } from "..";
@@ -36,6 +37,15 @@ const DIRECT_QUOTE_MODEL = {
// Room for the system prompt, the query, and the model's reasoning and answer.
const DIRECT_QUOTE_RESERVED_TOKENS = 16_384;
+const FREEFORM_GEMINI_MODEL = "gemini-3.5-flash-lite";
+
+// The Vertex provider falls back to the "google" key, so this covers both.
+const FREEFORM_PROVIDER_OPTIONS = {
+ google: {
+ thinkingConfig: { thinkingLevel: "minimal" },
+ } satisfies GoogleLanguageModelOptions,
+};
+
// o200k_base, which gpt-4o-mini uses and gpt-oss's o200k_harmony extends; a
// close enough estimate for Gemini to stay inside its much larger window.
const TOKENIZER_MODEL = "gpt-4o";
@@ -296,19 +306,16 @@ ${escapePromptTags(fitted.text)}
const modelChain = [
{
- name: "gemini-2.5-flash-lite",
- model: getModel(
- "gemini-2.5-flash-lite",
- hasVertex() ? "vertex" : "google",
- ),
+ name: FREEFORM_GEMINI_MODEL,
+ model: getModel(FREEFORM_GEMINI_MODEL, hasVertex() ? "vertex" : "google"),
},
{
name: "gpt-4o-mini",
model: getModel("gpt-4o-mini", "openai"),
},
{
- name: "gemini-2.5-flash-lite",
- model: getModel("gemini-2.5-flash-lite", "vertex"),
+ name: FREEFORM_GEMINI_MODEL,
+ model: getModel(FREEFORM_GEMINI_MODEL, "vertex"),
},
];
@@ -320,6 +327,7 @@ ${escapePromptTags(fitted.text)}
model,
system: querySystemPrompt,
prompt: queryPrompt,
+ providerOptions: FREEFORM_PROVIDER_OPTIONS,
experimental_telemetry: {
isEnabled: true,
functionId: "performQuery/freeform",