Refactor docs search provider exports

README.md

@@ -8,10 +8,11 @@ Shared docs tooling for Inth docs projects: React MDX rendering, MDX-to-markdown
 - `@inth/docs/remark`: remark plugins plus `defaultRemarkPlugins`
 - `@inth/docs/convert`: MDX-to-markdown conversion APIs
 - `@inth/docs/llm`: `llms.txt` and topic-scoped full-context generation
-- `@inth/docs/search`: edge-safe search runtime, content readers, guards, and rate limiter helpers
+- `@inth/docs/search`: search runtime, content readers, guards, and rate limiter helpers
 - `@inth/docs/search/node`: Node-only search index generation
-- `@inth/docs/search/ai`: AI SDK answer streaming helper
-- `@inth/docs/search/bash`: optional bash-tool docs inspection adapter
+- `@inth/docs/search/vercel`: Vercel AI Gateway / AI SDK answer streaming and bash tools
+- `@inth/docs/search/tanstack`: TanStack AI answer streaming and bash tools
+- `@inth/docs/search/cloudflare`: Cloudflare AI Gateway / Workers AI adapter helpers and bash tools
 - `@inth/docs/lint`: docs validation and the `inth-docs-lint` CLI
 
 ## Install
@@ -114,7 +115,7 @@ await generateDocsSearchFiles({
 });
 ```
 
-At runtime, query the generated JSON with `@inth/docs/search`. Add `@inth/docs/search/ai` only when a user explicitly asks for a source-grounded answer.
+At runtime, query the generated JSON with `@inth/docs/search`. Add a provider entrypoint such as `@inth/docs/search/vercel` only when a user explicitly asks for a source-grounded answer.
 
 ## Agent Docs
 

apps/docs-smoke/content/docs/guides/quickstart.mdx

@@ -234,7 +234,7 @@ export function search(query: string) {
 }
 ```
 
-If the app also supports AI answers, keep model calls behind an explicit user action and use `@inth/docs/search/ai` from a server route.
+If the app also supports AI answers, keep model calls behind an explicit user action and use a provider entrypoint such as `@inth/docs/search/vercel` from a server route.
 
 ## 7. Add Package Scripts
 

apps/docs-smoke/content/docs/index.mdx

@@ -32,19 +32,23 @@ description: "Developer reference for rendering MDX, converting docs, generating
     },
     "@inth/docs/search": {
       type: "runtime",
-      description: "Edge-safe search runtime, content readers, guards, and rate limiter helpers.",
+      description: "Search runtime, content readers, guards, and rate limiter helpers.",
     },
     "@inth/docs/search/node": {
       type: "build time",
       description: "Node-only `generateDocsSearchFiles`.",
     },
-    "@inth/docs/search/ai": {
+    "@inth/docs/search/vercel": {
       type: "optional runtime",
-      description: "AI SDK answer streaming helper.",
+      description: "Vercel AI Gateway / AI SDK answer streaming and bash tools.",
     },
-    "@inth/docs/search/bash": {
+    "@inth/docs/search/tanstack": {
       type: "optional runtime",
-      description: "bash-tool docs inspection adapter.",
+      description: "TanStack AI answer streaming and native docs bash tools.",
+    },
+    "@inth/docs/search/cloudflare": {
+      type: "optional runtime",
+      description: "Cloudflare AI Gateway / Workers AI adapters and docs bash tools.",
     },
     "@inth/docs/lint": {
       type: "build time",
@@ -63,7 +67,7 @@ description: "Developer reference for rendering MDX, converting docs, generating
     Use `@inth/docs/convert` with `@inth/docs/remark` to flatten authored MDX into markdown that works in LLM bundles and search indexes.
   </Step>
   <Step title="Generate and query search">
-    Generate static search JSON with `@inth/docs/search/node`, query it with `@inth/docs/search`, and add `@inth/docs/search/ai` only for explicit answer requests.
+    Generate static search JSON with `@inth/docs/search/node`, query it with `@inth/docs/search`, and add a provider entrypoint such as `@inth/docs/search/vercel` only for explicit answer requests.
   </Step>
 </Steps>
 

apps/docs-smoke/content/docs/search.mdx

@@ -11,6 +11,15 @@ description: "Generate static docs search data, query it at runtime, and stream
   Open [/search](/search) to test the generated index. Typing calls local search only. The `Ask` button is the only action that can call the model.
 </Callout>
 
+## Entrypoints
+
+| Use case | Import |
+| --- | --- |
+| Search, content reads, request guards, and rate limiting | `@inth/docs/search` |
+| Vercel AI Gateway / AI SDK answer + bash tools | `@inth/docs/search/vercel` |
+| TanStack AI answer + bash tools | `@inth/docs/search/tanstack` |
+| Cloudflare AI Gateway / Workers AI answer + bash tools | `@inth/docs/search/cloudflare` |
+
 ## Build The Index
 
 Use the Node-only entry point after MDX has been converted to markdown.
@@ -28,7 +37,7 @@ The generator writes `docs/search-index.json` for compact metadata and `docs/sea
 
 ## Runtime Search
 
-Use the edge-safe runtime with generated JSON.
+Use the provider-neutral runtime with generated JSON.
 
 ```ts
 import {
@@ -49,22 +58,59 @@ Results include heading paths, excerpts, `urlWithHash`, and `absoluteUrlWithHash
 
 ## Source-Grounded Answers
 
-Use `@inth/docs/search/ai` only when the user explicitly asks for an answer.
+Use a provider entrypoint only when the user explicitly asks for an answer. The demo app uses Vercel AI Gateway / AI SDK.
 
 ```ts
-import { streamDocsAnswer } from "@inth/docs/search/ai";
+import { streamDocsAnswer } from "@inth/docs/search/vercel";
 
 const { response, sources } = streamDocsAnswer({
   index,
   content,
   query,
-  model: process.env.DOCS_SEARCH_MODEL ?? "moonshotai/kimi-k2.6",
+  model: process.env.DOCS_SEARCH_MODEL ?? "openai/gpt-5.4-mini",
   productName: "@inth/docs",
 });
 ```
 
 The prompt includes retrieved docs context, asks for citations, treats docs text as untrusted reference content, and tells the model to say when the retrieved context is not enough.
 
+TanStack AI callers pass their adapter explicitly:
+
+```ts
+import { streamDocsAnswer } from "@inth/docs/search/tanstack";
+
+const { response, sources } = streamDocsAnswer({
+  index,
+  content,
+  query,
+  adapter,
+});
+```
+
+Cloudflare callers create an explicit TanStack-compatible adapter:
+
+```ts
+import {
+  createCloudflareDocsAdapter,
+  streamDocsAnswer,
+} from "@inth/docs/search/cloudflare";
+
+const adapter = createCloudflareDocsAdapter({
+  provider: "openai",
+  model: "gpt-4o",
+  options: {
+    binding: env.AI.gateway("docs-gateway"),
+  },
+});
+
+const { response, sources } = streamDocsAnswer({
+  index,
+  content,
+  query,
+  adapter,
+});
+```
+
 ## Abuse Protection
 
 <Steps>
@@ -81,14 +127,26 @@ The prompt includes retrieved docs context, asks for citations, treats docs text
 
 The demo uses an in-memory limiter for local smoke coverage. Production apps should adapt the `RateLimiter` interface to a shared store.
 
-## Bash Adapter
+## Bash Tools
 
-Use `@inth/docs/search/bash` when an AI SDK agent should inspect docs with safe shell commands.
+Use each provider entrypoint when an agent should inspect docs with safe shell commands. Tools are never created inside `streamDocsAnswer`; pass both tools and tool instructions explicitly.
 
 ```ts
-import { createDocsBashTool } from "@inth/docs/search/bash";
+import {
+  createDocsBashTool,
+  streamDocsAnswer,
+} from "@inth/docs/search/vercel";
 
 const { tools, instructions } = await createDocsBashTool(index, content);
+
+const { response } = streamDocsAnswer({
+  index,
+  content,
+  query,
+  model,
+  tools,
+  toolInstructions: instructions,
+});
 ```
 
 The adapter creates a read-only `/docs` filesystem for `just-bash` and wraps it with `bash-tool`. Agents can use commands such as `ls`, `cat`, `find`, `grep`, and `rg`; network commands and writes are disabled by default.

apps/docs-smoke/src/lib/docs.ts

@@ -79,22 +79,29 @@ export const packageSurfaces: PackageSurface[] = [
   {
     importPath: "@inth/docs/search",
     lifecycle: "runtime",
-    description: "Edge-safe search runtime, content readers, and guards.",
+    description:
+      "Search runtime, content readers, guards, and rate limiter helpers.",
   },
   {
     importPath: "@inth/docs/search/node",
     lifecycle: "build time",
     description: "Node-only `generateDocsSearchFiles`.",
   },
   {
-    importPath: "@inth/docs/search/ai",
+    importPath: "@inth/docs/search/vercel",
     lifecycle: "optional runtime",
-    description: "AI SDK answer streaming helper.",
+    description: "Vercel AI Gateway / AI SDK answer streaming and bash tools.",
   },
   {
-    importPath: "@inth/docs/search/bash",
+    importPath: "@inth/docs/search/tanstack",
     lifecycle: "optional runtime",
-    description: "bash-tool docs inspection adapter.",
+    description: "TanStack AI answer streaming and native docs bash tools.",
+  },
+  {
+    importPath: "@inth/docs/search/cloudflare",
+    lifecycle: "optional runtime",
+    description:
+      "Cloudflare AI Gateway / Workers AI adapter helpers and docs bash tools.",
   },
   {
     importPath: "@inth/docs/lint",

apps/docs-smoke/src/routes/api/docs/ask.ts

@@ -5,7 +5,7 @@ import {
   readJsonWithLimit,
   validateDocsQuery,
 } from "@inth/docs/search";
-import { streamDocsAnswer } from "@inth/docs/search/ai";
+import { streamDocsAnswer } from "@inth/docs/search/vercel";
 import { createFileRoute } from "@tanstack/react-router";
 import {
   docsSearchContent,
@@ -15,7 +15,7 @@ import {
   jsonResponse,
 } from "@/lib/search";
 
-const DEFAULT_MODEL = "moonshotai/kimi-k2.6";
+const DEFAULT_MODEL = "openai/gpt-5.4-mini";
 const DEFAULT_MAX_OUTPUT_TOKENS = 2000;
 const DEFAULT_TIMEOUT = {
   chunkMs: 15_000,

apps/docs-smoke/src/routes/playground.tsx

@@ -40,7 +40,7 @@ import { defaultRemarkPlugins, remarkInclude } from "@inth/docs/remark";`,
     summary:
       "Use the generated index for local search, then stream answers only when a user asks.",
     imports: `import { searchDocs } from "@inth/docs/search";
-import { streamDocsAnswer } from "@inth/docs/search/ai";`,
+import { streamDocsAnswer } from "@inth/docs/search/vercel";`,
     code: `const results = searchDocs(index, query, { content });
 
 const { response } = streamDocsAnswer({

apps/docs-smoke/tests/e2e/smoke.e2e.ts

@@ -20,7 +20,7 @@ test("home route renders the developer dashboard and package surfaces", async ({
 
   expect(html).toContain("Build docs with");
   expect(html).toContain("Implementation contract");
-  expect(html).toContain("@inth/docs/search/bash");
+  expect(html).toContain("@inth/docs/search/vercel");
 
   await page.goto("/", { waitUntil: "networkidle" });
   await expect(page.getByText(DASHBOARD_HEADING)).toBeVisible();
@@ -43,7 +43,7 @@ test("docs route renders package docs and extracted ExtractedTypeTable output",
   const html = await response.text();
 
   expect(html).toContain("@inth/docs");
-  expect(html).toContain("@inth/docs/search/bash");
+  expect(html).toContain("@inth/docs/search/cloudflare");
   expect(html).toContain("PipelineExampleOptions");
 
   await page.goto("/docs", { waitUntil: "networkidle" });