All demos require configuring an API key etc.
The demo consisted of loading a number (specifically, 11) popular Monty Python sketches in a database, and asking questions about them. The loading (ingestion) process was done ahead as it takes a long time.
The sketches were taken from
ibras.dk
and converted to WebVTT format
format with "voice annotations" (e.g. <v Shopkeeper>It's resting</v>)
with help from a popular LLM.
We have a driver program in the repo to ingest WebVTT files into a SQLite database.
This is tools/ingest_vtt.py. You run it as follows:
python tools/ingest_vtt.py FILE1.vtt ... FILEN.vtt -d mp.dbThe process took maybe 15 minutes for 11 sketches.
The sketches can now be queried by using another tool:
python tools/query.py -d mp.db(You just type questions and it prints answers.)
The demo consisted of loading a large number (around 500) email messages into a database, and querying the database about those messages. The loading (ingestion) process was done ahead as it takes a long time.
We used the Gmail API to download 550 messages from Guido's Gmail (details below).
Given a folder with *.eml files in MIME format, we ran our email
ingestion tool, tools/ingest_email.py. You run it as follows:
python tools/ingest_email.py -d gmail.db email-folder/You can also pass individual .eml files instead of a directory.
Use -v for verbose output.
Use --start-date and --stop-date to restrict ingestion to a date range with [start, stop):
# Ingest only January 2024 emails
python tools/ingest_email.py -d gmail.db email-folder/ \
--start-date 2024-01-01 --stop-date 2024-02-01These flags slice the input file list before any other filtering:
# Ingest only the first 20 files
python tools/ingest_email.py -d gmail.db email-folder/ --limit 20
# Skip the first 100 files, then process the next 50
python tools/ingest_email.py -d gmail.db email-folder/ \
--offset 100 --limit 50All four flags can be combined. The filter pipeline is: offset/limit → already-ingested → date range.
The process took over an hour for 500 messages. Moreover, it complained about nearly 10% of the messages due to timeouts or just overly large files. When an error occurs, the tool recovers and continues with the next file. Previously ingested emails are automatically skipped on subsequent runs.
We can then query the gmail.db database using the same query.py
tool that we used for the Monty Python demo:
python tools/query.py -d gmail.dbIn the tools/mail/ folder you'll find a tool named gmail_dump.py which
will download any number of messages (default 50) using the Gmail API.
In order to use the Gmail API, however, you have to create a (free)
Google Cloud app and configure it appropriately.
We created created an app in test mode at Google Cloud Console and gave it access to the Gmail API (I forget how exactly we did this part).
To create the needed client secret, we navigated to Client (side bar) and clicked on "+ Create Client" (in the row of actions at the top), selected "Desktop app", gave it a name, hit Create, scrolled down in the resulting dialog box, and hit "Download JSON". This produced a JSON file which should be copied into client_secret.json in the gmail folder. (The Cloud Console interface may look different for you.)
The first time you run the gmail_dump.py script, it will take you to a browser where you have to log in and agree to various warnings about using an app in test mode etc. The gmail_dump.py script then writes a file token.json and you're good for a week or so. When token.json expires, unfortunately you get a crash and you have to manually delete it to trigger the login flow again. (Sometimes starting a browser may fail, e.g. under WSL. Take the URL that's printed and manually go there.)
The rest of the email ingestion pipeline doesn't care where you got
your *.eml files from -- every email provider has its own quirks.
The podcast demo is actually the easiest to run:
The "database" is included in the repo as
tests/testdata/Episode_53_AdrianTchaikovsky_index*,
and this is in fact the default "database" used by tools/query.py
when no -d/--database flag is given.
This "database" indexes tests/testdata/Episode_53_AdrianTchaikovsky.txt.
It was created by a one-off script that invoked
src/typeagent/podcast/podcast_ingest/ingest_podcast()
and saved to two files by calling the .ingest() method on the
returned src/typeagent/podcasts/podcast/Podcast object.
Here's a brief sample session:
$ python tools/query.py
1.318s -- Using Azure OpenAI
0.054s -- Loading podcast from 'tests/testdata/Episode_53_AdrianTchaikovsky_index'
TypeAgent demo UI 0.2 (type 'q' to exit)
TypeAgent> What did Kevin say to Adrian about science fiction?
--------------------------------------------------
Kevin Scott expressed his admiration for Adrian Tchaikovsky as his favorite science fiction author. He mentioned that Adrian has a new trilogy called The Final Architecture, and Kevin is eagerly awaiting the third book, Lords of Uncreation, which he has had on preorder for months. Kevin praised Adrian for his impressive writing skills and his ability to produce large, interesting science fiction books at a rate of about one per year.
--------------------------------------------------
TypeAgent> How was Asimov mentioned.
--------------------------------------------------
Asimov was mentioned in the context of discussing the ethical and moral issues surrounding AI development. Adrian Tchaikovsky referenced Asimov's Laws of Robotics, noting that Asimov's stories often highlight the inadequacy of these laws in governing robots.
--------------------------------------------------
TypeAgent> q
$Enjoy exploring!