Merge pull request #1039 from syntaxfm/transcript-frontend

Transcript frontend + AI Show notes

package.json

@@ -53,7 +53,7 @@
 	},
 	"type": "module",
 	"engines": {
-		"node": ">=20.5.0",
+		"node": ">=18.0.0 <19.0.0 || =20.5.0",
 		"pnpm": ">=8"
 	},
 	"dependencies": {
@@ -67,8 +67,10 @@
 		"@types/emscripten": "^1.39.7",
 		"date-fns": "^2.30.0",
 		"flexsearch": "^0.7.31",
+		"gpt-3-encoder": "^1.1.4",
 		"gray-matter": "^4.0.3",
 		"media-chrome": "^1.2.2",
+		"openai": "^3.3.0",
 		"rehype-highlight": "^6.0.0",
 		"rehype-raw": "^6.1.1",
 		"rehype-stringify": "^9.0.3",

prisma/schema.prisma

@@ -174,7 +174,7 @@ model AiSummaryEntry {
 
 model AiTweet {
   id         Int        @id @default(autoincrement())
-  content    String     @db.VarChar(250)
+  content    String     @db.VarChar(350)
   showNote   Int
   aiShowNote AiShowNote @relation(fields: [showNote], references: [id])
 }

src/lib/player/CD.svelte

@@ -126,15 +126,15 @@ class="disc"
 		border-radius: 10px 10px 30px 30px;
 		background: linear-gradient(180deg, #000, #111);
 	}
-	
+
 	.case-glass {
 		border-radius: 10px 10px 30px 30px;
 		position: absolute;
 		top: 0;
 		width: 100%;
 		height: 95%;
 		background: linear-gradient(to right, #ffffff33, #ffffff66);
-		Z-index: 10;
+		z-index: 10;
 		left: 0;
 		box-shadow: inset 0 0 0px 4px #111, inset 0 0 0px 10px #222;
 		border: solid 2px #222;
@@ -151,7 +151,7 @@ class="disc"
 		border-bottom: solid 2px #111;
 		height: 60%;
 		background: radial-gradient(circle at center, #11111133,  #00000033), linear-gradient(140deg, #111, #333);
-		Z-index: 11;
+		z-index: 11;
 		left: 0;
 	}
 
@@ -161,7 +161,7 @@ class="disc"
 		width: 70px;
 		height: auto;
 	}
-	
+
 	.spinning {
 		animation: spin .5s linear infinite forwards;
 	}

src/server/ai/openai.ts

@@ -27,43 +27,31 @@ export const CONDENSE_THRESHOLD = 100;
 const configuration = new Configuration({
 	apiKey: process.env.OPENAI_API_KEY
 });
-// import { findSimilarUtterances } from './embeddings';
-import { error } from '@sveltejs/kit';
 import { AINoteSelect, AIPodcastSummaryResponse } from './queries';
 export const openai = new OpenAIApi(configuration);
 
-async function condense(transcript: string, show: TranscribedShow, inputTokensLength: number) {
-	console.log(`========== Condensing show ${show.number} ============`);
-	// See if we have a condensed transcript already saved to file
-	const path = `./deepgram/condensed-healed/${show.number}.json`;
-	const localExists = await exists(path);
-	// Save existing work to Memory
-	let existing: SlimUtterance[] = [];
-	if (localExists) {
-		const condensedFile = await readFile(path, 'utf-8');
-		console.log(`Using local condensed file for ${show.number}`);
-		existing = JSON.parse(condensedFile) as SlimUtterance[];
+export async function condense(
+	transcript: string,
+	show: TranscribedShow
+): Promise<SlimUtterance[]> {
+	// Figure out how large the input is
+	const inputTokensLength = encode(`${transcript} ${summarizePrompt}`).length;
+	// If its under the limit, return the transcript as is
+	if (inputTokensLength < TOKEN_INPUT_LIMIT) {
+		console.log(
+			`========== Skipping condensing show ${show.number} - Size is ${inputTokensLength} and acceptable ============`
+		);
+		return show.utterances;
 	}
-
+	console.log(`========== Condensing show ${show.number} ============`);
 	// Figure out how many hunks we need to split this string into
-	const size = encode(transcript).length;
-	console.log(`Transcript size: ${size}`);
 	console.log(`inputTokensLength size: ${inputTokensLength}`);
 	console.log(`Token input limit: ${TOKEN_INPUT_LIMIT}`);
 	const factorSmaller = 1 - TOKEN_INPUT_LIMIT / inputTokensLength;
 	console.log(`Factor smaller: ${factorSmaller}`);
 	// Split the transcript into hunks
 	const utteranceFuncs = show.utterances.map((utterance, index) => {
 		return async function getCondenseUtterance(): Promise<SlimUtterance> {
-			// Check if we already have this utterance condensed
-			const existingUtterance = existing.find(
-				(eu) => eu.utteranceIndex === utterance.utteranceIndex
-			);
-			if (existingUtterance) {
-				// console.log(utterance.showNumber, utterance.utteranceIndex, `Skipping OpenAI, we already have utterance condensed`);
-				return existingUtterance;
-			}
-
 			// Wait a random amount of time to avoid rate limiting. Between 0 and 10 seconds
 			const waitTime = Math.floor(Math.random() * 10000);
 			await wait(waitTime);
@@ -92,6 +80,7 @@ async function condense(transcript: string, show: TranscribedShow, inputTokensLe
 				console.dir(err.response.headers);
 			});
 			const condensed = completion?.data?.choices?.at(0)?.message?.content;
+			// Inject the condensed transcript into the utterance
 			if (condensed) {
 				utterance.condensedTranscript = condensed;
 			}
@@ -106,6 +95,7 @@ async function condense(transcript: string, show: TranscribedShow, inputTokensLe
 				)}% of original`
 			);
 			console.timeEnd(`Condensing ${index} of ${show.utterances.length}`);
+			// Return the modifined utterance
 			return utterance;
 		};
 	});
@@ -114,17 +104,12 @@ async function condense(transcript: string, show: TranscribedShow, inputTokensLe
 	console.log('Done condensing');
 	// Get the results
 	const utterances = utteranceResults
-		.filter((result): result is PromiseSettledResult<SlimUtterance> => {
+		.filter((result): result is PromiseFulfilledResult<SlimUtterance> => {
 			return result.status === 'fulfilled';
 		})
 		.map((result) => result.value);
 
 	console.log(`Saving condensed file for ${show.number}`);
-	// Write the results to a file
-	await writeFile(
-		`./deepgram/condensed-healed/${show.number}.json`,
-		JSON.stringify(utterances, null, 2)
-	);
 	return utterances;
 }
 
@@ -149,8 +134,14 @@ export async function generate_ai_notes(
 ) {
 	const slimUtterance = getSlimUtterances(show.transcript?.utterances || [], show.number);
 	const transcript = formatAsTranscript(slimUtterance);
-	// See if its over the token limit
-	// TODO: move to OpenAI's new functions API
+	// Condense
+	const slimUtterancesWithCondensed = await condense(transcript, {
+		name: show.title,
+		number: show.number,
+		utterances: slimUtterance
+	});
+	const condensedTranscript = formatAsTranscript(slimUtterancesWithCondensed);
+
 	const input: CreateChatCompletionRequest = {
 		model: MODEL,
 		messages: [
@@ -161,13 +152,14 @@ export async function generate_ai_notes(
 			},
 			{ role: 'user', content: `This episode is #${show.number} entitled ${show.title}` },
 			{ role: 'user', content: summarizePrompt },
-			{ role: 'user', content: transcript }
+			{ role: 'user', content: condensedTranscript }
 		]
 	};
 
 	console.log(`Creating AI notes for ${show.number}`);
 	const completion = await openai.createChatCompletion(input);
 	const maybeJSON = completion.data.choices.at(0)?.message?.content;
+	console.log(maybeJSON);
 	const parsed = JSON.parse(maybeJSON || '') as AIPodcastSummaryResponse;
 	return parsed;
 }

src/server/transcripts/deepgram.ts

@@ -1,4 +1,5 @@
-import { Deepgram } from '@deepgram/sdk';
+import DeepgramPkg from '@deepgram/sdk';
+const { Deepgram } = DeepgramPkg;
 import { prisma_client as prisma } from '../../hooks.server';
 import { error } from '@sveltejs/kit';
 import { keywords } from './fixes';
@@ -7,7 +8,7 @@ import { save_transcript_to_db } from './transcripts';
 import type { PrerecordedTranscriptionResponse } from '@deepgram/sdk/dist/types';
 const deepgramApiKey = process.env.DEEPGRAM_SECRET;
 if (!deepgramApiKey) {
-	console.log('Please set the DEEPGRAM_SECRET environment variable.');
+	console.error('Please set the DEEPGRAM_SECRET environment variable.');
 	process.exit(1);
 }
 

src/server/transcripts/utils.ts

@@ -79,12 +79,20 @@ export function formatTime(secs: number) {
 	return `${minutesString}:${secondsString}`;
 }
 
+function getSpeakerShortName(speaker: string | undefined) {
+	const shortForms = new Map([
+		['Scott Tolinski', 'Scott'],
+		['Wes Bos', 'Wes']
+	]);
+	return shortForms.get(speaker) || speaker;
+}
+
 export function formatAsTranscript(utterances: SlimUtterance[]) {
 	return utterances.reduce((acc, utterance) => {
 		// TODO: We might need to reinstate condensedTranscript here
 		const timestamp = formatTime(utterance.start);
-		return `${acc}\n${timestamp} ${utterance.speaker}:\n${
-			/*condensedTranscript || */ utterance.transcript
+		return `${acc}\n${timestamp} ${getSpeakerShortName(utterance.speaker)}:\n${
+			utterance.condensedTranscript || utterance.transcript
 		}\n`;
 	}, '');
 }