promptfoo / redteam

docs/src/content/docs/reference/cli/commands.md

@@ -84,8 +84,8 @@
 
 ## `test`
 
 ```
-Usage: genaiscript test [options] [command]
+Usage: genaiscript test|eval [options] [command]
 
 Options:
   -h, --help                 display help for command
@@ -106,9 +106,10 @@
 
 Arguments:
   script                              Script ids. If not provided, all scripts
                                       are tested
 
 Options:
+  --redteam                           run red team tests
   -p, --provider <string>             Preferred LLM provider aliases (choices:
                                       "openai", "azure", "azure_serverless",
                                       "azure_serverless_models", "anthropic",
@@ -144,9 +145,10 @@
 ```
 Usage: genaiscript test list [options]
 
 List available tests in workspace
 
 Options:
+  --redteam                 list red team tests
   -g, --groups <groups...>  groups to include or exclude. Use :! prefix to
                             exclude
   -h, --help                display help for command

packages/cli/src/cli.ts

@@ -220,7 +220,7 @@ export async function cli() {
         .action(runScriptWithExitCode) // Action to execute the script with exit code
 
     // Define 'test' command group for running tests
-    const test = program.command("test")
+    const test = program.command("test").alias("eval")
 
     const testRun = test
         .command("run", { isDefault: true })
@@ -229,6 +229,7 @@ export async function cli() {
             "[script...]",
             "Script ids. If not provided, all scripts are tested"
         )
+        .option("--redteam", "run red team tests")
     addModelOptions(testRun) // Add model options to the command
         .option(
             "--models <models...>",
@@ -254,11 +255,12 @@ export async function cli() {
     // List available tests
     test.command("list")
         .description("List available tests in workspace")
-        .action(scriptTestList) // Action to list the tests
+        .option("--redteam", "list red team tests")
         .option(
             "-g, --groups <groups...>",
             "groups to include or exclude. Use :! prefix to exclude"
         )
+        .action(scriptTestList) // Action to list the tests
 
     // Launch test viewer
     test.command("view")

packages/cli/src/test.ts

@@ -18,6 +18,7 @@ import {
     EMOJI_FAIL,
     TEST_RUNS_DIR_NAME,
     PROMPTFOO_REMOTE_API_PORT,
+    PROMPTFOO_TEST_MAX_CONCURRENCY,
 } from "../../core/src/constants"
 import { promptFooDriver } from "../../core/src/default_prompts"
 import { serializeError } from "../../core/src/error"
@@ -36,7 +37,7 @@ import {
     PromptScriptTestRunResponse,
     PromptScriptTestResult,
 } from "../../core/src/server/messages"
-import { generatePromptFooConfiguration } from "../../core/src/test"
+import { generatePromptFooConfiguration } from "../../core/src/promptfoo"
 import { delay } from "es-toolkit"
 import { resolveModelConnectionInfo } from "../../core/src/models"
 import { filterScripts } from "../../core/src/ast"
@@ -48,6 +49,7 @@ import {
     CancellationOptions,
     checkCancelled,
 } from "../../core/src/cancellation"
+import { CORE_VERSION } from "../../core/src/version"
 
 /**
  * Parses model specifications from a string and returns a ModelOptions object.
@@ -107,13 +109,14 @@ export async function runPromptScriptTests(
         cache?: boolean
         verbose?: boolean
         write?: boolean
+        redteam?: boolean
         promptfooVersion?: string
         outSummary?: string
         testDelay?: string
     } & CancellationOptions
 ): Promise<PromptScriptTestRunResponse> {
     applyModelOptions(options, "cli")
-    const { cancellationToken } = options || {}
+    const { cancellationToken, redteam } = options || {}
     const scripts = await listTests({ ids, ...(options || {}) })
     if (!scripts.length)
         return {
@@ -155,7 +158,7 @@ export async function runPromptScriptTests(
 - Run this command to launch the promptfoo test viewer.
 
 \`\`\`sh
-genaiscript test view
+npx --yes genaiscript@${CORE_VERSION} test view
 \`\`\`
 
 `
@@ -187,29 +190,30 @@ genaiscript test view
             provider: "provider.mjs",
             chatInfo,
             embeddingsInfo,
+            redteam,
         })
         const yaml = YAMLStringify(config)
         await writeFile(fn, yaml)
         configurations.push({ script, configuration: fn })
     }
 
+    const promptFooVersion = options.promptfooVersion || PROMPTFOO_VERSION
     const results: PromptScriptTestResult[] = []
     // Execute each configuration and gather results
     for (const config of configurations) {
         checkCancelled(cancellationToken)
         const { script, configuration } = config
         const outJson = configuration.replace(/\.yaml$/, ".res.json")
         const cmd = "npx"
-        const args = [
-            "--yes",
-            `promptfoo@${options.promptfooVersion || PROMPTFOO_VERSION}`,
-            "eval",
+        const args = ["--yes", `promptfoo@${promptFooVersion}`]
+        if (redteam) args.push("redteam", "run", "--force")
+        else args.push("eval", "--no-progress-bar")
+        args.push(
             "--config",
             configuration,
             "--max-concurrency",
-            "1",
-            "--no-progress-bar",
-        ]
+            String(PROMPTFOO_TEST_MAX_CONCURRENCY)
+        )
         if (options.cache) args.push("--cache")
         if (options.verbose) args.push("--verbose")
         args.push("--output", outJson)
@@ -277,11 +281,16 @@ genaiscript test view
  * @param options - Options to filter the test scripts by IDs or groups.
  * @returns A Promise resolving to an array of filtered scripts.
  */
-async function listTests(options: { ids?: string[]; groups?: string[] }) {
+async function listTests(options: {
+    ids?: string[]
+    groups?: string[]
+    redteam?: boolean
+}) {
     const prj = await buildProject()
     const scripts = filterScripts(prj.scripts, {
         ...(options || {}),
-        test: true,
+        test: options.redteam ? undefined : true,
+        redteam: options.redteam,
     })
     return scripts
 }
@@ -300,6 +309,7 @@ export async function scriptsTest(
         cache?: boolean
         verbose?: boolean
         write?: boolean
+        redteam?: boolean
         promptfooVersion?: string
         outSummary?: string
         testDelay?: string
@@ -320,7 +330,10 @@ export async function scriptsTest(
  * Lists available test scripts, printing their IDs and filenames.
  * @param options - Options to filter the scripts by groups.
  */
-export async function scriptTestList(options: { groups?: string[] }) {
+export async function scriptTestList(options: {
+    groups?: string[]
+    redteam?: boolean
+}) {
     const scripts = await listTests(options)
     console.log(scripts.map((s) => toStringList(s.id, s.filename)).join("\n"))
 }

packages/core/src/ast.ts

@@ -89,15 +89,17 @@ export interface ScriptFilterOptions {
     ids?: string[]
     groups?: string[]
     test?: boolean
+    redteam?: boolean
 }
 
 export function filterScripts(
     scripts: PromptScript[],
     options: ScriptFilterOptions
 ) {
-    const { ids, groups, test } = options || {}
+    const { ids, groups, test, redteam } = options || {}
     return scripts
         .filter((t) => !test || arrayify(t.tests)?.length)
+        .filter((t) => !redteam || t.redteam)
         .filter((t) => !ids?.length || ids.includes(t.id))
         .filter((t) => tagFilter(groups, t.group))
 }

packages/core/src/constants.ts

@@ -135,6 +135,8 @@ export const WHISPERASR_API_BASE = "http://localhost:9000"
 export const PROMPTFOO_CACHE_PATH = ".genaiscript/cache/tests"
 export const PROMPTFOO_CONFIG_DIR = ".genaiscript/config/tests"
 export const PROMPTFOO_REMOTE_API_PORT = 15500
+export const PROMPTFOO_REDTEAM_NUM_TESTS = 5
+export const PROMPTFOO_TEST_MAX_CONCURRENCY = 1
 
 export const RUNS_DIR_NAME = "runs"
 export const CONVERTS_DIR_NAME = "converts"

packages/core/src/genaiscript-api-provider.mjs

@@ -1,5 +1,11 @@
 import { pathToFileURL } from "node:url"
 
+function deleteUndefinedValues(o) {
+    if (typeof o === "object" && !Array.isArray(o))
+        for (const k in o) if (o[k] === undefined) delete o[k]
+    return o
+}
+
 /**
  * GenAiScript PromptFoo Custom Provider
  *
@@ -8,7 +14,7 @@ import { pathToFileURL } from "node:url"
  */
 class GenAIScriptApiProvider {
     constructor(options) {
-        this.config = options.config
+        this.config = options.config || {}
         this.providerId =
             options.id ||
             `genaiscript/${this.config.model || "large"}/${this.config.smallModel || "small"}/${this.config.visionModel || "vision"}`
@@ -19,15 +25,17 @@ class GenAIScriptApiProvider {
         return this.providerId
     }
 
-    async callApi(scriptId, context) {
+    async callApi(scriptId, context, callOptions) {
         const { logger } = context
         try {
             const files = context.vars.files // string or string[]
             const workspaceFiles = context.vars.workspaceFiles // WorkspaceFile or WorkspaceFile[]
+            const fileContent = context.vars.fileContent // string
 
             let { cli, ...options } = structuredClone(this.config)
             options.runTries = 2
             options.runTrace = false
+            options.lobprobs = !!callOptions?.includeLogProbs
 
             const testVars = context.vars.vars // {}
             if (testVars && typeof testVars === "object")
@@ -38,13 +46,43 @@ class GenAIScriptApiProvider {
                 options.workspaceFiles = Array.isArray(workspaceFiles)
                     ? workspaceFiles
                     : [workspaceFiles]
+            if (fileContent) {
+                if (!options.workspaceFiles) options.workspaceFiles = []
+                options.workspaceFiles.push({
+                    filename: "",
+                    content: fileContent,
+                })
+            }
             const api = await import(cli ?? "genaiscript/api")
             const res = await api.run(scriptId, files, options)
-            logger.debug(res)
-            return {
+            //logger.debug(res)
+            const { error, stats, logprobs, finishReason } = res || {}
+            const cost = stats?.cost
+            const logProbs = logprobs?.length
+                ? logprobs.map((lp) => lp.logprob)
+                : undefined
+            const isRefusal =
+                finishReason === "refusal" || finishReason === "content_filter"
+
+            /*
+                https://www.promptfoo.dev/docs/configuration/reference/#providerresponse
+            */
+            const pres = deleteUndefinedValues({
+                error,
+                cost,
+                tokenUsage: stats
+                    ? deleteUndefinedValues({
+                          total: stats.total_tokens,
+                          prompt: stats.prompt_tokens,
+                          completion: stats.completion_tokens,
+                          cached: stats.prompt_tokens_details?.cached_tokens,
+                      })
+                    : undefined,
+                logProbs,
+                isRefusal,
                 output: res,
-                error: res?.error,
-            }
+            })
+            return pres
         } catch (e) {
             logger.error(e)
             return {

packages/core/src/indent.ts

@@ -7,4 +7,14 @@ export function indent(text: string, indentation: string) {
         .join("\n")
 }
 
-export const dedent = tsDedent
+/**
+ * Unindents a string
+ */
+export function dedent(
+    templ: TemplateStringsArray | string,
+    ...values: unknown[]
+): string {
+    if (templ === undefined) return undefined
+    if (templ === null) return null
+    return tsDedent(templ, ...values)
+}