Transcript tools

This project is a showcase CLI tool to sales transcripts utilities with the help of GenAI.

Project structure:

• src/ - Source code of the project
• tests/ - Unit tests of the project
• i18n/ - Internationalization files for the transcripts and LLM prompts
• transcripts/ - The transcripts from either the example shared in the question or generated by the project
• mongodb/ - The MongoDB database volume would go there if you use the docker-compose file
• docker-compose.yml - The docker-compose file that can be used to spin up a MongoDB instance
• .env - The environment variables file. You can use it to set the environment variables for the project.

Project overview:

The project is a CLI tool that can be used to generate transcripts, summarize a transcript, and answer questions based on transcripts. Each of these functionalities is implemented in a separate file in the src/commands directory. The project uses the OpenAI API to perform these tasks. The project is structured in a way that it can be easily extended to add more functionalities. The chats with the LLM can be saved in a MongoDB database if the MONGODB_URI is provided in the .env file.

Commands overview:

The commands are built using the commander package, with the support of ora package to show the loading spinner while the command is waiting for the LLM to respond. tsx is used to execute the commands and mocha is used to run the tests. The commands are built in a way that you can choose to either having the necessary arguments in the command line when you start it or you can provide them later with a interactive prompt.

I18n support:

The project added support for the transcripts and questions in different languages. To be able to use this feature, you need to choose a model that has multilingual support. You can set the language in the command line with "-l" or "--lang" flag. The default language is English. The supported languages are English(en), Spanish(es), and French(fr). The supported foreign language is only used for the transcripts and the LLM chatting/response. The system messages are always in English.

Before running the project:

Before running the project, you need to install the dependencies. You can do that by running the following command:

npm install

You'll also need to set up the environment variables in the .env file. OPENAI_BASE_URL, OPENAI_API_KEY and OPENAI_MODEL are required, but you can also use a OpenAI compatible API like ollama. MONGODB_URI is optional, if it's provided, at the end of the ask command, the question and the answer will be saved in the database. TRANSCRIPT_OUTPUT_PATH is optional and defaults to transcripts/, it's the path where the generated transcripts will be saved. If you want to quickly spin up a MongoDB instance, you can use the docker-compose.yml file to do that. You can run the following command to start the MongoDB instance:

docker-compose up -d

How to run the project:

For the three parts of the project, each of them has its own subcommand. If you run the main command with

npm run start

It will show you the help message with the available subcommands. The help message is also available with the --help flag for both the main command and the subcommands.

---

Generating a transcript:

For the generation of the transcripts, you can run the following command:

npm run start generate

or simply

npm run generate
# Usage: transcripts-tools generate [options]
#
# Generate a sales call transcript
#
# Options:
#   -f, --file <string>  save transcript to the file (default: "")
#   -l, --lang <string>  language in use (choices: "en", "es", "fr", default: "en")
#   -h, --help           display help for command

This would start the generation of the transcript in the language specified. While it's being generated, you'll see a preview of it in the console. The final transcript will be post processed to clean it up. If you specified the file name in the command line like this:

npm run generate -- -f transcript1.txt

The transcript will be saved in the transcripts/ directory with the name transcript1.txt. If you didn't specify the file name, the command will ask you if you want to save it or not.

---

Summarizing a transcript:

For the summarization of the transcripts, you can run the following command:

npm run summarize
# Usage: transcripts-tools summarize [options]
#
# Summarize a sales call transcript
#
# Options:
#   -f, --file <string>  read transcript from the file (default: "")
#   -l, --lang <string>  language in use (choices: "en", "es", "fr", default: "en")
#   -h, --help           display help for command

Same as the generate command, you can choose to specify the path to the transcript file from the command line or there will be an interactive prompt to ask you for it. The summary will be shown in the console in the language specified.

---

Asking questions:

For asking questions based on the transcripts, you can run the following command:

npm run ask
# Usage: transcripts-tools ask [options]
#
# Ask questions about a sales call transcript
#
# Options:
#   -f, --file <string>      read transcript from the file (default: "")
#   -q, --question <string>  question about the transcript (default: "")
#   -l, --lang <string>      language in use (choices: "en", "es", "fr", default: "en")
#   -h, --help               display help for command

You can specify the path to the transcript file and the question from the command line or there will be an interactive prompt to ask you for them. If you provided a initial question in the command line, it'll be used to ask the LLM. The answer will be shown in the console. After each question, there would be a prompt to ask if you want to ask another question or not. If no more questions are entered, the command would try to save the questions and answers in the database if the MONGODB_URI is provided and then exit.

Results:

Sharing a few results from the commands:

Generate transcript in French:

npm run generate -- -l fr

> [email protected] generate
> tsx src generate -l fr

[ASSISTANT] Below is the preview of the transcript:

00:00:00 Alex (google.ai) : Bonjour Aisha.
00:00:02 Aisha (ibm.watson) : Salut Alex, comment allez-vous ?
00:00:05 Alex (google.ai) : Très bien, merci. Nous avons besoin de 8000 TPU supplémentaires pour nos derniers modèles d'apprentissage en profondeur. Êtes-vous en mesure d'honorer cette demande ?
00:00:06 Aisha (ibm.watson) : Nous comprenons l'importance des TPU dans votre travail, Alex, mais actuellement nous ne pouvons pas fournir le nombre exact de 8000 que vous requérez. Cependant, si vous êtes flexible avec ce chiffre, nous pourrions trouver une solution plus réaliste comme 4500.
00:00:10 Alex (google.ai) : Je comprends votre position, Aisha, mais nos modèles ont besoin de toute la puissance de calcul disponible pour atteindre leur plein potentiel. Serait-il possible d'augmenter ce chiffre d'ici quelques mois ? Nous avons des délais serrés à respecter.
00:00:20 Aisha (ibm.watson) : Je compatis avec votre situation. Malheureusement, l'allocation de ressources est actuellement fixée à un niveau inférieur. Si vous êtes ouvert à une solution progressive, nous pourrions augmenter le nombre de TPU à mesure que nous avons des disponibilités supplémentaires. Cependant, au départ, je peux vous offrir 4500 TPU.
00:00:35 Alex (google.ai) : Cela devient un peu problématique pour nous, Aisha. Nous pourrions peut-être rencontrer à mi-chemin avec quelque chose comme 6000 TPU ? Cela nous donnerait plus de puissance de calcul pour les mois à venir et nous aiderait à respecter nos délais.
00:00:50 Aisha (ibm.watson) : J'apprécie votre souplesse, Alex. Je vais discuter avec notre équipe technique et voir si une augmentation progressive du nombre de TPU peut être possible dans un avenir proche. Nous allons vous contacter pour discuter des détails plus approfondis.
00:01:00 Alex (google.ai) : Cela me semble faisable, Aisha. Merci beaucoup pour votre temps et votre considération. Nous attendons votre appel.
00:01:10 Aisha (ibm.watson) : De rien, Alex. Nous allons trouver une solution qui convienne à tous les deux. Talk soon.
00:01:15 Alex (google.ai) : À bientôt, Aisha.


[ASSISTANT] Above is the generated transcript.
[SYSTEM] Do you want to save the transcript? Type 'no' to skip:

Summarize the transcript:

npm run summarize -- -l es -f transcripts/testes1.txt

> [email protected] summarize
> tsx src summarize -l es -f transcripts/testes1.txt

[ASSISTANT] Below is the summary of the transcript:

El resumen de la conversación es el siguiente:

Comenzó con Sam (de openai.com) llamando a Maria (de ibm.com) para hablar sobre su acuerdo de procesamiento cognitivo. Sam quería aumentar su capacidad en un 15% debido al crecimiento de su empresa. Sin embargo, Maria respondió que actualmente estaban ajustados en cuanto a capacidad, y solo podían ofrecer un aumento del 10%. Sam pidió considerar un aumento mayor, como el 12% o el 15%, pero eso no fue posible según Maria.

Después de negociar, llegaron a un acuerdo: comenzarán con un aumento del 10% durante seis meses, luego reevaluarán las capacidades para Betrachtel ein mögliches erhöhtes zur Verfügung stehen. Sam aceptó agradeciendo la comprensión de Maria.

Transparency on AI Usage:

The GenAI is used in generating the LLM prompts and transcript examples. More details in src/i18n/en.ts and other files in the src/i18n/ directory. This document is written with the help of GenAI. Other than that, the project doesn't use any GenAI or similar services on the creating the code.