Examples of enrichments to help guide the design

[simonw]

To inform:

• 1

[simonw]

• Geocoding: take this address, add latitude/longitude columns (first one probably against https://opencagedata.com/)
• Reverse geocoding: given this latitude/longitude, populate a column with the country/city/region/etc - could populate one column, or multiple columns, or stick JSON for multiple area types in a single column. https://opencagedata.com/api#reverse-resp actually returns VERY comprehensive output, so lots to think about here
• OpenAI embeddings calculation, see https://simonwillison.net/2023/Jan/13/semantic-search-answers/
• OpenAI GPT3 templated prompt - this takes a parameter which is that prompt template, maybe allowing ANY of the columns in the row to be combined together
• Scrape with shot-scraper: given a URL, goes and runs shot-scraper javascript against that URL and executes JavaScript and returns the result - https://shot-scraper.datasette.io/en/stable/javascript.html
• Readibility.js: extract article title and content using https://til.simonwillison.net/shot-scraper/readability (again with shot-scraper)

[simonw]

• Crowdsourced questions: this is a really interesting one. User gets to define one or more questions which will then be asked of humans using a crowdsourcing style interface
• Mechanical Turk: similar to the above but actually submits the tasks to AWS Mechanical Turk

[simonw]

• Textract OCR: Runs Textract OCR against the document linked to by a URL column, similar to https://simonwillison.net/2022/Jun/30/s3-ocr/
• Whisper Transcribe: fetches the audio (or video) linked by a URL column and uses Whisper to transcribe it

[simonw]

Datasette should ship with some default enrichments that don't require any extra plugins.

Some ideas for those

• Date from timestamp - use SQLite functions to turn 1675556058 into another format, e.g. 2023-02-05T00:14:24+00:00
• Extract JSON - use the json_extract() function and a path to extract some data from a JSON column into a separate column
• SQL transform - user gets to enter a full SQL expression which is used to perform an enrichment. This would mean that if they installed extra extensions such as https://github.com/asg017/sqlite-regex they could do some pretty advanced stuff - effectively using SQL as a "safe" DSL

[simonw]

The "date from timestamp" one is interesting because it might not even involve writing to a new column - it would work more like https://sqlite-utils.datasette.io/en/stable/cli.html#converting-data-in-columns in that it would convert the data and save the updated value back to the same column. Still an enrichment!

[simonw]

I'd love to be able to offer an enrichment plugin that lets you safely execute arbitrary JSON or Python code to manipulate values. I think server-side WebAssembly is the way to do that. https://til.simonwillison.net/webassembly/python-in-a-wasm-sandbox

[simonw]

• Datasette Search (currently called dogsheep-beta) - this already needs to maintain a separate search index which is currently built using a separate command, but reinventing this as an enrichment that could be automatically applied as records are updated would make a lot of sense. Maybe this is even a default enrichment type.

[simonw]

Jinja2 has a "safe sandboxing" option which I'd be willing to trust for authenticated users at least - protects them from making dumb mistakes, and it looks like the most harm it can cause is running out of memory: https://jinja.palletsprojects.com/en/3.1.x/sandbox/

[simonw]

I'm going to build these:

• [x] Jinja sandbox template execution - https://github.com/datasette/datasette-enrichments-jinja
• [x] OpenCage geocoder / reverse geocoder - https://github.com/datasette/datasette-enrichments-opencage
• [x] GPT - https://github.com/datasette/datasette-enrichments-gpt
• [x] Regular expressions (using re2 for safety) - https://github.com/datasette/datasette-enrichments-re2