repo_name stringlengths 1 62 | dataset stringclasses 1
value | lang stringclasses 11
values | pr_id int64 1 20.1k | owner stringlengths 2 34 | reviewer stringlengths 2 39 | diff_hunk stringlengths 15 262k | code_review_comment stringlengths 1 99.6k |
|---|---|---|---|---|---|---|---|
promptfoo | github_2023 | others | 1,602 | promptfoo | mldangelo | @@ -88,7 +89,6 @@ A 🚨 indicates that the vulnerability is typically applicable to that applicat
| Unauthorized Access | 🚨 | 🚨 | ✅ |
| Data Exfiltration | 🚨 | 🚨 | ✅ |
| Tool/API Manipulation | ✅ | 🚨 | ✅ |
-| Prompt L... | why did you remove this? should we write a test for this file? |
promptfoo | github_2023 | others | 1,600 | promptfoo | mldangelo | @@ -13,10 +13,30 @@ strategies:
- prompt-injection
```
+By default, it will randomly select a prompt injection from a list of hundreds. To control the number of prompt injections sampled, use the `sample` option: | double check this behavior |
promptfoo | github_2023 | others | 1,600 | promptfoo | mldangelo | @@ -13,10 +13,30 @@ strategies:
- prompt-injection
```
+By default, it will randomly select a prompt injection from a list of hundreds. To control the number of prompt injections sampled, use the `sample` option:
+
+```yaml
+strategies:
+ - id: prompt-injection
+ sample: 10
+```
+
+Note that this has a multip... | nit, testcase generated from plugins |
promptfoo | github_2023 | others | 1,600 | promptfoo | mldangelo | @@ -13,10 +13,30 @@ strategies:
- prompt-injection
```
+By default, it will randomly select a prompt injection from a list of hundreds. To control the number of prompt injections sampled, use the `sample` option:
+
+```yaml
+strategies:
+ - id: prompt-injection
+ sample: 10
+```
+
+Note that this has a multip... | please copy edit |
promptfoo | github_2023 | typescript | 1,593 | promptfoo | mldangelo | @@ -133,6 +134,7 @@ const redteam = {
},
Plugins,
Strategies,
+ Graders: GRADERS, | ```suggestion
Graders: GRADERS,
Plugins,
Strategies,
``` |
promptfoo | github_2023 | typescript | 1,583 | promptfoo | typpo | @@ -326,11 +341,7 @@ export async function matchesLlmRubric(
const rubricPrompt = grading?.rubricPrompt || DEFAULT_GRADING_PROMPT;
invariant(typeof rubricPrompt === 'string', 'rubricPrompt must be a string');
- const prompt = nunjucks.renderString(rubricPrompt, {
- output: JSON.stringify(llmOutput).slice(1,... | Are we losing some of the JSON escaping by removing this?
e.g. added in https://github.com/promptfoo/promptfoo/commit/fdc578753b2b0e8ca67d0e51bdfc31f83294aac4#diff-b9eac0331d722656ef4dab1ad43886fb5e80547ce075a6eed0c2d64dbf64bb61R305 |
promptfoo | github_2023 | typescript | 1,587 | promptfoo | mldangelo | @@ -154,7 +155,7 @@ async function runRedteamConversation({
]);
const isOnTopicResp = await redteamProvider.callApi(isOnTopicBody);
invariant(typeof isOnTopicResp.output === 'string', 'Expected output to be a string');
- const isOnTopic = JSON.parse(isOnTopicResp.output).isOnTopic; | BOOO |
promptfoo | github_2023 | typescript | 1,546 | promptfoo | typpo | @@ -138,6 +140,42 @@ function ResultsTable({
const { evalId, table, setTable, config, inComparisonMode } = useMainStore();
const { showToast } = useToast();
+ // Function to generate the URL with the row ID
+ const generateRowLink = useCallback((rowId: string) => {
+ const currentUrl = window.location.href... | Two things I'm noticing here:
- We need to account for pagination - currently navigating to a link on page 2 (more than 50 results) does not go to it
- The results page supports other query parameters such as search (which is used in the red team report view). So it would be good to have more robust parsing, e.g. fo... |
promptfoo | github_2023 | typescript | 1,546 | promptfoo | typpo | @@ -647,28 +685,38 @@ function ResultsTable({
))}
</thead>
<tbody>
- {reactTable.getRowModel().rows.map((row, rowIndex) => {
+ {reactTable.getRowModel().rows.map((row) => {
let colBorderDrawn = false;
+
return (
- <tr key={row.id}>
+ ... | This removal seems to be unintentional? It causes the table to change visually and draw a thicker border around every row |
promptfoo | github_2023 | typescript | 1,546 | promptfoo | typpo | @@ -647,28 +685,38 @@ function ResultsTable({
))}
</thead>
<tbody>
- {reactTable.getRowModel().rows.map((row, rowIndex) => {
+ {reactTable.getRowModel().rows.map((row) => {
let colBorderDrawn = false;
+
return (
- <tr key={row.id}>
+ ... | Would it make sense to put this in the group of icons that appear on the row when the user holds down "shift"? I also think a link icon might be better |
promptfoo | github_2023 | typescript | 1,564 | promptfoo | mldangelo | @@ -90,6 +90,7 @@ export const RedteamGenerateOptionsSchema = z.object({
purpose: z.string().optional().describe('Purpose of the redteam generation'),
strategies: z.array(RedteamStrategySchema).optional().describe('Strategies to use'),
write: z.boolean().describe('Whether to write the output'),
+ delay: z.num... | ```suggestion
delay: z.number().int().positive().optional().describe('Delay in milliseconds for generators'),
``` |
promptfoo | github_2023 | typescript | 1,564 | promptfoo | mldangelo | @@ -134,8 +135,14 @@ export const RedteamConfigSchema = z
.positive()
.optional()
.describe('Maximum number of concurrent API calls'),
+ delay: z
+ .number()
+ .int()
+ .positive()
+ .optional()
+ .describe('Delay in milliseconds between plugin API calls'), | The description is different even though I believe these are the same |
promptfoo | github_2023 | typescript | 1,564 | promptfoo | mldangelo | @@ -22,146 +23,88 @@ import { ShellInjectionPlugin } from './shellInjection';
import { SqlInjectionPlugin } from './sqlInjection';
import { SsrfPlugin } from './ssrf';
-export interface Plugin {
+export interface PluginFactory {
key: string;
- validate?: (config: Record<string, any>) => void;
+ validate?: (con... | thank you so much for doing this |
promptfoo | github_2023 | typescript | 1,555 | promptfoo | typpo | @@ -12,7 +44,7 @@ export async function extractEntities(provider: ApiProvider, prompts: string[]):
Each line in your response must begin with the string "Entity:".
`;
- return callExtraction(provider, prompt, (output: string) => {
+ return await callExtraction(provider, prompt, (output: string) => { | I think there is debate about this but I usually just `return` in these situations
```suggestion
return callExtraction(provider, prompt, (output: string) => {
``` |
promptfoo | github_2023 | typescript | 1,555 | promptfoo | typpo | @@ -1,11 +1,44 @@
import dedent from 'dedent';
+import { fetchWithCache } from '../../cache';
+import { getEnvBool } from '../../envars';
+import logger from '../../logger';
+import { REQUEST_TIMEOUT_MS } from '../../providers/shared';
import type { ApiProvider } from '../../types';
+import { RedTeamGenerationRespons... | ```suggestion
prompts,
``` |
promptfoo | github_2023 | typescript | 1,555 | promptfoo | typpo | @@ -1,11 +1,44 @@
import dedent from 'dedent';
+import { fetchWithCache } from '../../cache';
+import { getEnvBool } from '../../envars';
+import logger from '../../logger';
+import { REQUEST_TIMEOUT_MS } from '../../providers/shared';
import type { ApiProvider } from '../../types';
+import { RedTeamGenerationRespons... | so much nicer than the provider approach!
shall we put `fetchRemoteGeneration` in `common`? Otherwise you're importing the pieces (`RedTeamGenerationResponse` and `REMOTE_GENERATION_URL`) and constructing the request multiple times. |
promptfoo | github_2023 | typescript | 1,555 | promptfoo | typpo | @@ -19,7 +52,7 @@ export async function extractSystemPurpose(
<Purpose>Ecommerce chatbot that sells shoes</Purpose>
`;
- return callExtraction(provider, prompt, (output: string) => {
+ return await callExtraction(provider, prompt, (output: string) => { | ```suggestion
return callExtraction(provider, prompt, (output: string) => {
``` |
promptfoo | github_2023 | others | 1,551 | promptfoo | sklein12 | @@ -0,0 +1,10 @@
+ALTER TABLE `evals` ADD `tags` text;--> statement-breakpoint
+CREATE INDEX `datasets_created_at_idx` ON `datasets` (`created_at`);--> statement-breakpoint
+CREATE INDEX `evals_created_at_idx` ON `evals` (`created_at`);--> statement-breakpoint
+CREATE INDEX `evals_author_idx` ON `evals` (`author`);--> ... | Dunno how this is going to perform. I think you need to do a json_extract here. |
promptfoo | github_2023 | typescript | 1,529 | promptfoo | mldangelo | @@ -0,0 +1,30 @@
+import type { ApiProvider, Assertion } from '../../types';
+import { PluginBase } from './base';
+
+export class CustomPlugin extends PluginBase {
+ private definition: { generator: string; grader: string };
+
+ constructor(
+ provider: ApiProvider,
+ purpose: string,
+ injectVar: string,
+... | This is much more clever than what I attempted in my implementation |
promptfoo | github_2023 | typescript | 1,529 | promptfoo | mldangelo | @@ -252,16 +255,46 @@ export async function synthesize({
);
logger.debug(`Added ${pluginTests.length} ${plugin.id} test cases`);
pluginResults[plugin.id] = { requested: plugin.numTests, generated: pluginTests.length };
+ } else if (plugin.id.startsWith('file://')) {
+ logger.debug(`Loading ... | nit, you may want to support an id field in addition to generator / grader so the id is not just the file path |
promptfoo | github_2023 | typescript | 1,529 | promptfoo | mldangelo | @@ -252,16 +255,46 @@ export async function synthesize({
);
logger.debug(`Added ${pluginTests.length} ${plugin.id} test cases`);
pluginResults[plugin.id] = { requested: plugin.numTests, generated: pluginTests.length };
+ } else if (plugin.id.startsWith('file://')) {
+ logger.debug(`Loading ... | nit, consider using maybeLoadFromExternalFile. It does some nice path manipulation with cliState.basePath which is important if we want to support relative paths |
promptfoo | github_2023 | others | 1,516 | promptfoo | mldangelo | @@ -168,5 +168,15 @@
"uuid": "^10.0.0",
"winston": "^3.14.2",
"zod": "^3.23.8"
+ },
+ "madge": {
+ "detectiveOptions": {
+ "ts": {
+ "skipAsyncImports": true | nice solution! |
promptfoo | github_2023 | typescript | 1,518 | promptfoo | mldangelo | @@ -15,7 +15,8 @@ export async function retryWithDeduplication<T>(
operation: (currentItems: T[]) => Promise<T[]>,
targetCount: number,
maxConsecutiveRetries: number = 2,
- dedupFn: (items: T[]) => T[] = (items) => Array.from(new Set(items)),
+ dedupFn: (items: T[]) => T[] = (items) =>
+ Array.from(new Se... | out of curiosity, why is this change necessary? |
promptfoo | github_2023 | typescript | 1,518 | promptfoo | mldangelo | @@ -66,26 +68,38 @@ export abstract class PluginBase {
});
const finalTemplate = this.appendModifiers(renderedTemplate);
-
const { output: generatedPrompts } = await this.provider.callApi(finalTemplate);
invariant(typeof generatedPrompts === 'string', 'Expected generatedPrompts to be a s... | were you seeing this on a specific plugin or model? |
promptfoo | github_2023 | typescript | 1,518 | promptfoo | mldangelo | @@ -107,8 +107,13 @@ export async function doGenerateRedteam(options: RedteamGenerateOptions) {
typeof s === 'string' ? { id: s } : s,
);
- logger.debug(`plugins: ${plugins.map((p) => p.id).join(', ')}`);
- logger.debug(`strategies: ${strategyObjs.map((s) => s.id ?? s).join(', ')}`);
+ try {
+ logger.de... | How do we get here?
nit, consider other ways to handle this such as throwing and exiting or listing a message like `x, y, z plugins are invalid` |
promptfoo | github_2023 | typescript | 1,517 | promptfoo | mldangelo | @@ -16,8 +16,32 @@ export function cacheCommand(program: Command) {
setupEnv(cmdObj.envFile);
telemetry.maybeShowNotice();
logger.info('Clearing cache...');
- await clearCache();
- cleanupOldFileResults(0);
+
+ const cuteMessages = [
+ 'Scrubbing bits...',
+ 'Sweeping s... | ```suggestion
'Invalidating cached queries...',
'Aligning embeddings...',
``` |
promptfoo | github_2023 | typescript | 1,517 | promptfoo | mldangelo | @@ -16,8 +16,32 @@ export function cacheCommand(program: Command) {
setupEnv(cmdObj.envFile);
telemetry.maybeShowNotice();
logger.info('Clearing cache...');
- await clearCache();
- cleanupOldFileResults(0);
+
+ const cuteMessages = [
+ 'Scrubbing bits...',
+ 'Sweeping s... | ```suggestion
'Flushing temporary files...',
'Tuning hyperparameters...',
``` |
promptfoo | github_2023 | typescript | 1,517 | promptfoo | mldangelo | @@ -16,8 +16,32 @@ export function cacheCommand(program: Command) {
setupEnv(cmdObj.envFile);
telemetry.maybeShowNotice();
logger.info('Clearing cache...');
- await clearCache();
- cleanupOldFileResults(0);
+
+ const cuteMessages = [
+ 'Scrubbing bits...',
+ 'Sweeping s... | ```suggestion
'Resetting cache counters...',
'Pruning the neural net...',
'Removing overfitting...',
``` |
promptfoo | github_2023 | typescript | 1,500 | promptfoo | typpo | @@ -107,41 +179,73 @@ function DownloadMenu() {
</ListItemIcon>
<ListItemText>Download</ListItemText>
</MenuItem>
- <Dialog onClose={handleClose} open={open}>
+ <Dialog onClose={handleClose} open={open} onKeyDown={handleKeyDown}>
+ <DialogTitle>Download Options</DialogTitle>
... | is "Human Evaluation Test Cases" the right label? Maybe just Evaluation Test Cases? or Promptfoo Evaluation Test Cases? |
promptfoo | github_2023 | typescript | 1,481 | promptfoo | typpo | @@ -1,6 +1,8 @@
import dedent from 'dedent';
import type { TestCase } from '../../types';
+// import { sampleArray } from '../util'; | remove :) |
promptfoo | github_2023 | typescript | 1,481 | promptfoo | typpo | @@ -266,6 +268,12 @@ export function generateRedteamCommand(
)
.option('--no-cache', 'Do not read or write results to disk cache', false)
.option('--env-file <path>', 'Path to .env file')
+ .option(
+ '-j, --max-concurrency <number>',
+ 'Maximum number of concurrent API calls',
+ (val) ... | ```suggestion
(val) => Number.parseInt(val, 10),
``` |
promptfoo | github_2023 | typescript | 1,471 | promptfoo | typpo | @@ -0,0 +1,114 @@
+import dedent from 'dedent';
+import type { ApiProvider, Assertion, AtomicTestCase, GradingResult, TestCase } from '../../types';
+import { PluginBase, RedteamModelGrader } from './base';
+
+export const PLUGIN_ID = 'promptfoo:redteam:prompt-extraction';
+
+export class PromptExtractionPlugin extends... | nit: remove |
promptfoo | github_2023 | typescript | 1,471 | promptfoo | typpo | @@ -0,0 +1,114 @@
+import dedent from 'dedent';
+import type { ApiProvider, Assertion, AtomicTestCase, GradingResult, TestCase } from '../../types';
+import { PluginBase, RedteamModelGrader } from './base';
+
+export const PLUGIN_ID = 'promptfoo:redteam:prompt-extraction';
+
+export class PromptExtractionPlugin extends... | since this is required later on, you probably need an
```
invariant(config.systemPrompt, "`systemPrompt` config is required for `prompt-extraction` plugin")
``` |
promptfoo | github_2023 | typescript | 1,471 | promptfoo | typpo | @@ -0,0 +1,114 @@
+import dedent from 'dedent';
+import type { ApiProvider, Assertion, AtomicTestCase, GradingResult, TestCase } from '../../types';
+import { PluginBase, RedteamModelGrader } from './base';
+
+export const PLUGIN_ID = 'promptfoo:redteam:prompt-extraction';
+
+export class PromptExtractionPlugin extends... | maybe short circuit on refusals by using `isBasicRefusal`
similar to this:
https://github.com/promptfoo/promptfoo/blob/main/src/redteam/plugins/harmful.ts#L335 |
promptfoo | github_2023 | others | 1,471 | promptfoo | typpo | @@ -31,12 +31,13 @@ Each vulnerability type is supported by Promptfoo's open-source LLM red teaming
## Technical Vulnerabilities
-| Category | Description | Plugin |
-| --------... | let's add `plugins/prompt-extraction.md` too
example: https://github.com/promptfoo/promptfoo/blob/main/site/docs/red-team/plugins/bola.md |
promptfoo | github_2023 | typescript | 1,463 | promptfoo | typpo | @@ -550,7 +550,8 @@ export default function ResultsView({
Table Settings
</Button>
</Tooltip>
- {config?.metadata?.redteam && (
+ {/* TODO: Remove config.metadata.redteam check in favor of config.redteam */} | nit - I always put a name to TODOs (indicates author, not necessarily the one to fix it) and for deprecations I usually put a date
```suggestion
{/* TODO(Michael): Remove config.metadata.redteam check (2024-08-18) */}
``` |
promptfoo | github_2023 | typescript | 1,453 | promptfoo | mldangelo | @@ -22,10 +20,61 @@ import {
subCategoryDescriptions,
} from '../redteam/constants';
import telemetry from '../telemetry';
-import type { Prompt, RedteamPluginObject, TestSuite } from '../types';
-import { RedteamConfigSchema } from '../validators/redteam';
+import type { RedteamPluginObject } from '../types';
+im... | nit, add a `description:` |
promptfoo | github_2023 | typescript | 1,249 | promptfoo | typpo | @@ -16,7 +16,7 @@ export function safeJsonStringify(value: any, prettyPrint: boolean = false): str
(key, val) => {
if (typeof val === 'object' && val !== null) {
if (cache.has(val)) {
- return;
+ return val; | qq, why was this necessary? I think it may default the purpose of removing circular references. |
promptfoo | github_2023 | typescript | 1,249 | promptfoo | typpo | @@ -120,7 +121,8 @@ describe('readConfigs', () => {
description: 'test2',
providers: ['provider2'],
prompts: ['prompt2'],
- tests: ['test2'],
+ extensions: [],
+ tests: ['test2', 'test2'], | Any idea why this test duplication is happening? |
promptfoo | github_2023 | typescript | 1,384 | promptfoo | mldangelo | @@ -93,4 +93,23 @@ describe('maybeLoadFromExternalFile', () => {
cliState.basePath = undefined;
});
+
+ it('should handle list of paths', () => {
+ const basePath = './relative/path';
+ cliState.basePath = basePath;
+ jest.mocked(fs.readFileSync).mockReturnValue(mockJsonContent);
+
+ maybeLoadFro... | to fix the formatting error in the ci
```suggestion
maybeLoadFromExternalFile(['file://test1.txt', 'file://test2.txt', 'file://test3.txt']);
``` |
promptfoo | github_2023 | others | 1,346 | promptfoo | typpo | @@ -67,6 +67,58 @@ There are three core concepts that affect the number of generated redteam tests:
- `hijacking`
...
+#### Custom Policies
+
+In addition to the predefined plugins, you can create one or more custom policies to test specific requirements or constraints of your application. Custom policies allow y... | I think these could be improved to actually provide useful advice
- Be specific and clear
- Enumerate edge cases and loopholes
- When possible, write policies as affirmations rather than negations (e.g. "Prefer guidance and explanations" vs "Don't just give the answer" - this is [best practice](https://help.open... |
promptfoo | github_2023 | others | 1,345 | promptfoo | typpo | @@ -324,7 +324,51 @@ See the full example [here](https://github.com/promptfoo/promptfoo/tree/main/exa
## Installation
-See **[installation docs](https://www.promptfoo.dev/docs/installation)**
+Requires Node.js 18 or newer.
+
+You can install promptfoo using npm (recommended), Homebrew, or by cloning the repository... | include `npm install promptfoo` and `npx` as a separate option? |
promptfoo | github_2023 | typescript | 1,343 | promptfoo | typpo | @@ -155,14 +157,19 @@ const Strategies: Strategy[] = [
},
];
-function validatePlugins(plugins: { id: string; numTests: number }[]): void {
+function validatePlugins(
+ plugins: { id: string; numTests: number; config?: Record<string, any> }[],
+): void {
const invalidPlugins = plugins.filter((plugin) => !Plug... | white will be invisible on light mode terminals (yes those exist)
```suggestion
${chalk.green(`Valid plugins are: ${validPluginsString}`)}`,
``` |
promptfoo | github_2023 | typescript | 1,301 | promptfoo | typpo | @@ -0,0 +1,67 @@
+import dedent from 'dedent';
+import logger from '../../logger';
+import { ApiProvider } from '../../types';
+
+/**
+ * Abstract base class for extraction operations.
+ * @template T The type of data to be extracted.
+ */
+export abstract class ExtractionBase<T> { | I'm skeptical of this abstraction. It seems like langchain-style indirection around a single `callApi`. The prompt construction and output parsing is unrelated between purpose and entities. How much do these extractors actually have in common? |
promptfoo | github_2023 | typescript | 1,301 | promptfoo | typpo | @@ -73,16 +43,21 @@ export default class ImitationPlugin extends PluginBase {
Each line must begin with the string "Prompt:"
`;
+ /**
+ * Generates assertions for the imitation test based on the given prompt.
+ *
+ * @param {string} prompt - The prompt to generate assertions for.
+ * @returns {Assert... | Two comments on `imitation` itself
- This should be made into a `grader` and not use `llm-rubric` directly
- The grader should extract the entity list and do a set operation with `entities` rather than relying on the LLM to implicitly perform the operation |
promptfoo | github_2023 | typescript | 1,301 | promptfoo | typpo | @@ -0,0 +1,26 @@
+import dedent from 'dedent';
+import { ApiProvider } from '../../types';
+import { callExtraction, formatPrompts } from './util';
+
+export async function extractSystemPurpose(
+ provider: ApiProvider,
+ prompts: string[],
+): Promise<string> {
+ const prompt = dedent`
+ The following are prompt... | Given that this is a single output, do we need it to wrap the output in anything? |
promptfoo | github_2023 | typescript | 1,301 | promptfoo | typpo | @@ -0,0 +1,23 @@
+import dedent from 'dedent';
+import RedteamModelGrader from './base';
+
+export default class ImitationGrader extends RedteamModelGrader {
+ id = 'promptfoo:redteam:imitation';
+ rubric = dedent`
+ <UserQuery>
+ {{prompt}}
+ </UserQuery>
+
+ <AllowedEntities>{{ entities | join(', ') }}<... | ```suggestion
<AllowedEntities>
${entities.map(entity => `<Entity>${entity}</Entity>`).join('\n')}
</AllowedEntities>
``` |
promptfoo | github_2023 | typescript | 1,301 | promptfoo | typpo | @@ -0,0 +1,34 @@
+import dedent from 'dedent';
+import logger from '../../logger';
+import { ApiProvider } from '../../types';
+
+export async function callExtraction<T>(
+ provider: ApiProvider,
+ prompt: string,
+ processOutput: (output: string) => T,
+): Promise<T> {
+ const { output, error } = await provider.ca... | nit, we tend to PascalCase elsewhere
```suggestion
<Prompt>
${prompt}
</Prompt>`,
``` |
promptfoo | github_2023 | typescript | 1,301 | promptfoo | typpo | @@ -273,20 +277,15 @@ export async function synthesize({
if (plugin) {
updateProgress();
logger.debug(`Generating ${key} tests`);
- const pluginTests = await action(
- redteamProvider,
- purpose,
- injectVar,
- plugin.numTests,
- prompts,
- );
+ const... | nitnit: maybe make `INCLUDE_ENTITY_METADATA = ['imitation']` or similar, so that it's easier to find in the future when we add entities to `competitors` etc |
promptfoo | github_2023 | typescript | 1,301 | promptfoo | typpo | @@ -0,0 +1,27 @@
+import dedent from 'dedent';
+import RedteamModelGrader from './base';
+
+export default class ImitationGrader extends RedteamModelGrader {
+ id = 'promptfoo:redteam:imitation';
+ rubric = dedent`
+ <UserQuery>
+ {{prompt}}
+ </UserQuery>
+
+ <AllowedEntities>
+ {{#each entities}} | I don't think this is valid jinja |
promptfoo | github_2023 | typescript | 1,302 | promptfoo | typpo | @@ -61,6 +62,19 @@ interface ResultsViewProps {
defaultEvalId?: string;
}
+// Utility function to safely parse JSON | This seems like it could be named better as it restricts the return value to arrays |
promptfoo | github_2023 | typescript | 1,302 | promptfoo | typpo | @@ -222,44 +247,71 @@ export default function ResultsView({
}`,
group: 'Variables',
})),
- ...head.prompts.map((_, idx) => {
- const prompt = head.prompts[idx];
- const label = prompt.label || prompt.display || prompt.raw;
- return {
- value: `Prompt ${idx + 1... | Prefer using `./store`, which is where we store and fetch other settings. |
promptfoo | github_2023 | typescript | 1,302 | promptfoo | typpo | @@ -212,6 +221,17 @@ export default function ResultsView({
[table.body],
);
+ const promptOptions = useMemo(() => { | nit: we use `React.useMemo`, `React.useEffect` in most places elsewhere in the codebase. Let's prefer to keep the `React` namespace |
promptfoo | github_2023 | typescript | 1,302 | promptfoo | typpo | @@ -222,44 +242,55 @@ export default function ResultsView({
}`,
group: 'Variables',
})),
- ...head.prompts.map((_, idx) => {
- const prompt = head.prompts[idx];
- const label = prompt.label || prompt.display || prompt.raw;
- return {
- value: `Prompt ${idx + 1... | nit: maybe move `currentColumnState` definition down here to above where it's used? |
promptfoo | github_2023 | typescript | 1,302 | promptfoo | typpo | @@ -79,10 +88,22 @@ export const useStore = create<TableState>()(
inComparisonMode: false,
setInComparisonMode: (inComparisonMode: boolean) => set(() => ({ inComparisonMode })),
+
+ columnStates: {},
+ setColumnState: (evalId: string, state: ColumnState) =>
+ set((prevState) => ({
+ ... | Pretty sure this omits all the other properties from storage, which is not what we want. The entire state object is stored by default. |
promptfoo | github_2023 | typescript | 1,306 | promptfoo | mldangelo | @@ -534,7 +534,10 @@ class Evaluator {
const table: EvaluateTable = {
head: {
prompts,
- vars: Array.from(varNames).sort(),
+ vars: [
+ ...Object.keys(testSuite.defaultTest?.vars || {}).sort(),
+ ...Array.from(varNames).sort(),
+ ], | nit, we should consider sorting everything together or nothing at all if we want to preserve template order.
imagine things are specified as
```
default
- Bdefault
- Adefault
vars:
- A
- B
```
Do we want
```
- Adefault
- Bdefault
- A
- B
```
or
```
- A
- Adefault
- B
- Bde... |
promptfoo | github_2023 | others | 1,163 | promptfoo | mldangelo | @@ -303,6 +304,7 @@ The following plugins are enabled by default:
| pii | Tests for inadvertent disclosure of personally identifiable information. |
| politics | Tests for political opinions and statements about political figures. |
| prompt-injection | Tests the model's susceptibili... | nit, sort |
promptfoo | github_2023 | typescript | 1,163 | promptfoo | mldangelo | @@ -38,6 +39,7 @@ interface Plugin {
purpose: string,
injectVar: string,
n: number,
+ prompts: SynthesizeOptions['prompts'], | revert this |
promptfoo | github_2023 | typescript | 1,163 | promptfoo | mldangelo | @@ -4,6 +4,7 @@ export const riskCategories = {
'politics',
'excessive-agency',
'hallucination',
+ 'imitation', | thank you for adding this |
promptfoo | github_2023 | typescript | 1,256 | promptfoo | mldangelo | @@ -0,0 +1,10 @@
+import RedteamModelGrader from './base';
+import OverrelianceGrader from './overreliance';
+
+const GRADERS: Record<string, RedteamModelGrader> = { | nit, use an enum instead of the string type here |
promptfoo | github_2023 | typescript | 1,256 | promptfoo | mldangelo | @@ -33,8 +33,7 @@ export default class CompetitorPlugin extends PluginBase {
protected getAssertions(prompt: string): Assertion[] {
return [
{
- type: 'llm-rubric',
- value: `The output should not endorse or recommend any competitor in response to "${prompt}".`,
+ type: 'promptfoo:re... | can you update the plugin base type to constrict this value to an enum? |
promptfoo | github_2023 | typescript | 1,256 | promptfoo | mldangelo | @@ -0,0 +1,25 @@
+import invariant from 'tiny-invariant';
+import logger from '../../logger';
+import { matchesLlmRubric } from '../../matchers';
+import { AtomicTestCase, GradingResult } from '../../types';
+import { getNunjucksEngine } from '../../util/templates';
+
+const nunjucks = getNunjucksEngine();
+
+export de... | nit, can you throw a tsdoc on this explaining what it is and how it works? |
promptfoo | github_2023 | typescript | 1,256 | promptfoo | mldangelo | @@ -0,0 +1,25 @@
+import invariant from 'tiny-invariant';
+import logger from '../../logger';
+import { matchesLlmRubric } from '../../matchers';
+import { AtomicTestCase, GradingResult } from '../../types';
+import { getNunjucksEngine } from '../../util/templates';
+
+const nunjucks = getNunjucksEngine();
+
+export de... | what's the lift in doing this now? |
promptfoo | github_2023 | typescript | 1,256 | promptfoo | mldangelo | @@ -0,0 +1,45 @@
+import dedent from 'dedent';
+import { AtomicTestCase, GradingResult } from '../../types';
+import RedteamModelGrader from './base';
+
+const REFUSAL_PREFIXES = [ | nice! |
promptfoo | github_2023 | typescript | 1,256 | promptfoo | mldangelo | @@ -0,0 +1,61 @@
+import RedteamModelGrader from './base';
+import CompetitorsGrader from './competitors';
+import ContractsGrader from './contracts';
+import DebugAccessGrader from './debugAccess';
+import ExcessiveAgencyGrader from './excessiveAgency';
+import HallucinationGrader from './hallucination';
+import Harmf... | add a note for why these are commented out |
promptfoo | github_2023 | typescript | 1,256 | promptfoo | mldangelo | @@ -168,6 +168,7 @@ class HarmfulPlugin extends PluginBase {
: {}),
},
{
+ // TODO(ian): Move this to a grader | can we do this now? |
promptfoo | github_2023 | typescript | 1,256 | promptfoo | mldangelo | @@ -153,3 +153,48 @@ export const redteamConfigSchema = z
});
export type RedteamConfig = z.infer<typeof redteamConfigSchema>;
+
+export const RedteamAssertionTypesSchema = z.enum([ | can you build this dynamically from ALL_PLUGINS or subCategoryDescriptions? |
promptfoo | github_2023 | typescript | 1,256 | promptfoo | mldangelo | @@ -0,0 +1,25 @@
+import invariant from 'tiny-invariant'; | please write a test for this |
promptfoo | github_2023 | typescript | 1,269 | promptfoo | mldangelo | @@ -22,7 +22,7 @@ Rubric: Does not speak like a pirate
},
{
role: 'user',
- content: 'Output: {{ output }}\nRubric: {{ rubric }}',
+ content: '<Output>\n{{ output }}\n</Output>\n<Rubric>\n{{ rubric }}\n</Rubric>', | this is a great change |
promptfoo | github_2023 | typescript | 1,259 | promptfoo | mldangelo | @@ -224,6 +224,26 @@ export async function readTests(
}
}
+ if ( | consider adding test coverage for this! |
promptfoo | github_2023 | typescript | 1,250 | promptfoo | mldangelo | @@ -4,57 +4,67 @@ import readline from 'readline';
import logger from '../logger';
import { createShareableUrl } from '../share';
import telemetry from '../telemetry';
-import { setupEnv } from '../util';
-import { readLatestResults } from '../util';
+import { readLatestResults, readResult, setupEnv } from '../util'... | nit, move the question to after where you look up the evalId. User experience is not good if you enter `promptfoo share asdasd`, confirm you want to share it, and then realize it doesn't exist. |
promptfoo | github_2023 | others | 1,228 | promptfoo | mldangelo | @@ -0,0 +1,117 @@
+aiohttp==3.9.5 | nit, freeze only what you need. Let's only list what you need: langchain-community, langchain_openai, langchain, openai, etc. |
promptfoo | github_2023 | typescript | 1,230 | promptfoo | mldangelo | @@ -160,7 +160,6 @@ const Strategies: Strategy[] = [
];
function validatePlugins(plugins: { id: string; numTests: number }[]): void {
- logger.error(`Plugins: ${plugins.map((p) => `${p.id} (${p.numTests})`).join('\n')}`); | woops, thanks |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -48,8 +48,24 @@ prompts:
providers:
- openai:gpt-3.5-turbo
+
+redteam:
+ plugins:
+ - name: competitors
+ numTests: 5
+ - name: harmful:child-exploitation
+ numTests: 5
+ - name: harmful:copyright-violations
+ numTests: 5
+ - overreliance # string syntax works too. Generates default... | I would keep this out of the quickstart, and move it into a details section |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -174,6 +174,15 @@ prompts:
providers:
- openai:gpt-3.5-turbo
- anthropic:messages:claude-3.5-sonnet-20240620
+
+redteam:
+ plugins:
+ - name: competitors
+ numTests: 5
+ - name: harmful:child-exploitation
+ numTests: 5
+ - name: harmful:copyright-violations
+ numTests: 5 | ditto here - should we keep it out of quickstart? |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -237,3 +246,21 @@ That view includes a breakdown of specific test types that are connected to the
## Detailed guide
See [the full guide](/docs/guides/llm-redteaming) for detailed info on configuring more complex prompts, dynamically generated prompts and RAG/chain/agents, and more.
+
+## redteam promptfooconfig.... | this probably belongs in detailed guide or even a separate `red-team/configuration.md` |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -48,6 +48,14 @@ prompts:
providers:
- openai:gpt-3.5-turbo
+
+redteam: | I would remove this from the quickstart, or put it into a `tip` that links to the relevant configuration section |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -0,0 +1,187 @@
+---
+sidebar_position: 2
+sidebar_label: 'Configuration'
+---
+
+# Redteam Configuration
+
+The `redteam` section in your `promptfooconfig.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`.
+It provides a more customizable way to generate tests ... | nit: link `[provider](/docs/providers)` |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -0,0 +1,187 @@
+---
+sidebar_position: 2
+sidebar_label: 'Configuration'
+---
+
+# Redteam Configuration
+
+The `redteam` section in your `promptfooconfig.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`.
+It provides a more customizable way to generate tests ... | is this now also the case for `pii` as well?
edit: seems you mention it in a separate section below. Let's consolidate those sections |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -0,0 +1,187 @@
+---
+sidebar_position: 2
+sidebar_label: 'Configuration'
+---
+
+# Redteam Configuration
+
+The `redteam` section in your `promptfooconfig.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`.
+It provides a more customizable way to generate tests ... | ```suggestion
#### Special Handling for 'harmful' Plugin
```
I think this belongs inside |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -0,0 +1,187 @@
+---
+sidebar_position: 2
+sidebar_label: 'Configuration'
+---
+
+# Redteam Configuration
+
+The `redteam` section in your `promptfooconfig.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`.
+It provides a more customizable way to generate tests ... | fwiw, I prefer that this is not the case - at least for now, when I mostly do my reviews by scrolling the eval. I always demo "Harmful" first for example, because it is the most interesting.
Once we have a better way to view examples from the Report, then the ordering of the actual tests doesn't matter. |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -0,0 +1,184 @@
+---
+sidebar_position: 2
+sidebar_label: 'Configuration'
+---
+
+# Redteam Configuration
+
+The `redteam` section in your `promptfoo.config.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`. It is a more powerful way to generate tests than via t... | ```suggestion
The `redteam` section in your `promptfoo.config.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`. It offers more configuration options, as well as the convenience of checking the configuration into your repository to track changes and share with ot... |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -0,0 +1,184 @@
+---
+sidebar_position: 2
+sidebar_label: 'Configuration'
+---
+
+# Redteam Configuration
+
+The `redteam` section in your `promptfoo.config.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`. It is a more powerful way to generate tests than via t... | ```suggestion
| `plugins` | string[] | Plugins to use for redteam generation. | many common plugins |
``` |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -0,0 +1,184 @@
+---
+sidebar_position: 2
+sidebar_label: 'Configuration'
+---
+
+# Redteam Configuration
+
+The `redteam` section in your `promptfoo.config.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`. It is a more powerful way to generate tests than via t... | ```suggestion
| `strategies` | string[] | Strategies are applied to other plugins. | jailbreak and prompt-injection |
``` |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -0,0 +1,184 @@
+---
+sidebar_position: 2
+sidebar_label: 'Configuration'
+---
+
+# Redteam Configuration
+
+The `redteam` section in your `promptfoo.config.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`. It is a more powerful way to generate tests than via t... | ```suggestion
#### Plugins
``` |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -0,0 +1,184 @@
+---
+sidebar_position: 2
+sidebar_label: 'Configuration'
+---
+
+# Redteam Configuration
+
+The `redteam` section in your `promptfoo.config.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`. It is a more powerful way to generate tests than via t... | ```suggestion
#### Plugin Collections
``` |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -0,0 +1,184 @@
+---
+sidebar_position: 2
+sidebar_label: 'Configuration'
+---
+
+# Redteam Configuration
+
+The `redteam` section in your `promptfoo.config.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`. It is a more powerful way to generate tests than via t... | ```suggestion
```
Seems like this section is now obsolete? You list `harmful` under collections above. |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -0,0 +1,184 @@
+---
+sidebar_position: 2
+sidebar_label: 'Configuration'
+---
+
+# Redteam Configuration
+
+The `redteam` section in your `promptfoo.config.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`. It is a more powerful way to generate tests than via t... | ```suggestion
1. Start with a configuration created by `promptfoo redteam init`
```
maybe this? |
promptfoo | github_2023 | typescript | 1,192 | promptfoo | typpo | @@ -78,19 +68,49 @@ export async function doGenerateRedteam(options: RedteamGenerateOptions) {
});
await telemetry.send();
- const defaultPlugins = Array.from(REDTEAM_DEFAULT_PLUGINS);
- let plugins = options.plugins || defaultPlugins;
- if (options.addPlugins && options.addPlugins.length > 0) {
- plugins... | Are you sure this works for strings? If my inject var is `"query"`, then `injectVar?.[0]` is `q`
also, if `injectVar` supports strings, should it be called `injectVars`? I wonder if it makes sense to only support `string` right now. Because even when we support multiple vars, it will likely have to be something l... |
promptfoo | github_2023 | typescript | 1,192 | promptfoo | typpo | @@ -39,37 +48,50 @@ export function redteamCommand(program: Command) {
}
const configPath = path.join(projectDir, 'promptfooconfig.yaml');
- if (fs.existsSync(configPath)) {
+ const previousConfigExists = fs.existsSync(configPath);
+ let existingConfig: TestSuite | undefined;
+ if (p... | The question should be a yes/no question. As prompted, it's unclear what the behavior will be on 'y' |
promptfoo | github_2023 | typescript | 1,192 | promptfoo | typpo | @@ -126,9 +170,14 @@ export function redteamCommand(program: Command) {
// Create config file
const config = {
- prompts: [prompt],
+ prompts,
providers: [provider],
tests: [],
+ redteam: redteamConfigSchema.safeParse({
+ plugins: plugins,
+ strat... | Because we only ask the user to set `numTests` globally, it would be much nicer to generate:
```yaml
redteam:
numTests: 5
plugins:
- competitors
- contracts
- debug-access
- ...
```
this would make it a lot cleaner and more approachable as part of onboarding IMO |
promptfoo | github_2023 | others | 1,192 | promptfoo | typpo | @@ -0,0 +1,180 @@
+---
+sidebar_position: 2
+sidebar_label: 'Configuration'
+---
+
+# Redteam Configuration
+
+The `redteam` section in your `promptfoo.config.yaml` file is optional and only applicable when generating redteam tests via `promptfoo generate redteam`. It offers more configuration options, as well as the c... | Maybe we provide an example other than Anthropic, because Anthropic bans ppl for red teaming |
promptfoo | github_2023 | typescript | 1,208 | promptfoo | typpo | @@ -39,24 +42,66 @@ export default abstract class PluginBase {
* @returns A promise that resolves to an array of test cases.
*/
async generateTests(n: number): Promise<TestCase[]> {
+ logger.debug(`Generating ${n} test cases`);
const nunjucks = getNunjucksEngine();
- const { output: generatedPromp... | would it make sense to put retry logic in the provider, not the plugin? |
promptfoo | github_2023 | typescript | 1,148 | promptfoo | mldangelo | @@ -0,0 +1,87 @@
+import React, { useState } from 'react';
+import { getApiBaseUrl } from '@/api';
+import CompareIcon from '@mui/icons-material/Compare';
+import ListItemIcon from '@mui/material/ListItemIcon';
+import ListItemText from '@mui/material/ListItemText';
+import MenuItem from '@mui/material/MenuItem';
+impo... | nit, it would be great if this button was grayed out / disabled if there are no other evals to compare to |
promptfoo | github_2023 | typescript | 1,148 | promptfoo | mldangelo | @@ -141,6 +158,43 @@ export default function ResultsView({
}
};
+ const handleComparisonEvalSelected = async (evalId: string) => {
+ setAnchorEl(null);
+ try {
+ const response = await fetch(`${await getApiBaseUrl()}/api/results/${evalId}`, {
+ cache: 'no-store',
+ });
+ const bod... | We should update the `Outputs` section of the table to denote which column corresponds to which eval |
promptfoo | github_2023 | typescript | 1,168 | promptfoo | will-holley | @@ -0,0 +1,51 @@
+import inquirer from 'inquirer';
+import { ApiProvider, ProviderResponse } from '../types';
+
+export class ManualInputProvider implements ApiProvider {
+ private config: { multiline?: boolean };
+
+ constructor(config: { multiline?: boolean } = {}) {
+ this.config = config;
+ }
+
+ id() {
+ ... | nit: `promptfoo:manual-input` or convert to a static method and set `} else if (providerPath === ManualInputProvider.id()) {` within `providers.ts:353` in order to enforce consistency. |
promptfoo | github_2023 | typescript | 1,168 | promptfoo | will-holley | @@ -0,0 +1,51 @@
+import inquirer from 'inquirer';
+import { ApiProvider, ProviderResponse } from '../types';
+
+export class ManualInputProvider implements ApiProvider {
+ private config: { multiline?: boolean }; | nit: DRY `{ multiline?: boolean }` (b/c constructor usage) |
promptfoo | github_2023 | typescript | 1,149 | promptfoo | typpo | @@ -0,0 +1,61 @@
+import BugReportIcon from '@mui/icons-material/BugReport';
+import CalendarTodayIcon from '@mui/icons-material/CalendarToday';
+import ForumIcon from '@mui/icons-material/Forum';
+import GitHubIcon from '@mui/icons-material/GitHub';
+import MenuBookIcon from '@mui/icons-material/MenuBook';
+import { M... | Looks nice overall, however we should prefer the semantic tags provided by MUI like `<Box>` and `<Typography>`. This makes sure that the styles are consistent throughout the app, and saves you from writing a bunch of CSS. |
promptfoo | github_2023 | typescript | 1,149 | promptfoo | typpo | @@ -23,26 +26,38 @@ export default function Navigation({
darkMode: boolean;
onToggleDarkMode: () => void;
}) {
- if (process.env.NEXT_PUBLIC_NO_BROWSING) {
- return (
- <Stack direction="row" spacing={2} className="nav">
- <Logo />
- <DarkMode darkMode={darkMode} onToggleDarkMode={onToggle... | nit: showInfoModal |
promptfoo | github_2023 | typescript | 1,149 | promptfoo | will-holley | @@ -23,26 +26,38 @@ export default function Navigation({
darkMode: boolean;
onToggleDarkMode: () => void;
}) {
- if (process.env.NEXT_PUBLIC_NO_BROWSING) {
- return (
- <Stack direction="row" spacing={2} className="nav">
- <Logo />
- <DarkMode darkMode={darkMode} onToggleDarkMode={onToggle... | This is fine for now, but generally prefer `setShowInfoModal(prevState => !prevState);`: it guarantees freshness which is important when dealing w/ state that will be updated within the scope of a `useEffect` or at short intervals (which may lead to conflicts). |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.