Extending my personal infrastructure with a data source [see within blog graph]

Featuring Roam Research

I have a frustration around the fact that our personal data is scattered, siloed and incredibly hard to use for any purposes, not anticipated by the original developers.

One of my big efforts is unifying all of this scattered data into a system that makes it easier to oversee, combine, interact with and build new tools around it.

In this post I want to illustrate how to integrate a new data source in the system, with Roam Research data as an example.

I've chosen Roam for this post because:

  • it's a cool, useful and rich data source
  • it hasn't been integrated before in my system, so it's easy to follow and document all the steps necessary

Throughout the post, I will link to my previous writing on building the , so I won't go into explaining every tiny detail and technical rationales here. But if you have some questions, feel free to ask anyway!

1 What is Roam Research?

Roam is a new note taking/thinking tool. If you're an org-mode user, you can think of it as having a web app with a directory of org files with the main focus (as I see it) on:

You can learn more about Roam here, which is a welcome page and a sandbox at the same time (very cool!).

If you want to get a glimpse of Org-mode workflow without having to use Emacs, Roam is probably the tool you should try.

So, let's say you've been using Roam and want to use its data outside of the web app, for some other purpose. What do you do?

For this demo, I'll be using a public covid19 Roam database.

2 Export raw data

The first step is figuring out how to get the raw data.

Roam doesn't have a public API yet, but they let you export a JSON from the web interface, which is good enough for us. How to proceed with this knowledge? I apply the principles from "Building data liberation infrastructure":

  • use an existing library to fetch the data

    It is counterproductive to reimplement this all over again. Luckily, there is one, MatthieuBizien/roam-to-git! It's very decent and does all the heavy lifting like browser automation.

    The only thing I had to do is implementing a wrapper to call it and extract the JSON from the archive.

  • keep the raw data intact

    I simply store the JSON without any changes on my filesystem. There is not much point trying to squeeze inherently hierarchical Roam data into a database. If performance ever becomes an issue, a transparent caching layer can always be added.

    Roam is under active development, so it's likely that the export format will change. It would be pretty sad if it breaks exports, and keeping the data unchanged means I always have a recent version on my disk.

3 Automate export

Ok, now we can get a snapshot of Roam data with a single script run: ./roamresearch.py > export.json. What's next?

The idea is that we run it periodically, for example, few times a day. It does mean that the data on your disk is lagging behind slightly, but that gives you 90% of the benefits without all the downsides of implementing a realtime synchronization. You also need an API for two-way synchronization anyway (unless you want to poll the service like crazy).

Running it manually is not much fun, and there are several tools that can aid us in it. I've done an overview in this post, and personally, I'm using an experimental tool, dron, which is basically a nice Systemd frontend. But it really doesn't matter what exactly to use, so the simplest is a good old cron job:

00 05 * * * /soft/export-scripts/roamresearch.py > /exports/roamresearch/data.json

This will export your data every day at 5:00 and save it to a JSON file. The file will be used by other tools and scripts to get the most recent snapshot of the data.

  • For me the command looks a bit more sophisticated:

    /soft/arctee/arctee.py /exports/roamresearch/roamresearch_{utcnow}.json -- /soft/export-scripts/roamresearch.py
    

    This keeps every export along with a timestamp, ensures atomic writing to minimize concurrency issues and retries automatically. I elaborate more on it here, but it's not super crucial for Roam, more of a nice-to-have thing.

Once you automate JSON exports, you also get Roam data backups for free (you back up your disk, right?).

There is a problem though. Working with JSON is okay for a one-off thing, but if you want to truly benefit from using the data for different applications, you better off having some abstract interface for it. While Roam's JSON export is fairly reasonable (unlike many other services), it's pretty annoying to figure out the schema

Take a look at the raw data. It's clear, that you'll need (at least) to:

  • convert integer timestamps into datetime objects
  • keep track of optional and required fields
  • deconstruct the hierarchy

This is annoying, boring and error-prone to do all over again. We need some central way to encapsulate and adapt raw JSON data for easier programmatic use. Say hello to the data access layer!

4 Adapt the data

The data access layer is responsible for deserializing and converting raw (i.e. JSON/sqlite) format into abstract objects, easy to work in your code.

Normally I'd try to use existing data bindings, but Roam is pretty new and I haven't found any, so I implemented my own. The data adapter is integrated into HPI package, which is a central entry point for all of my data access.

The whole thing is contained within a single pull request and takes about 100 LOC.

If you browse through the code you can also see the importance of raw data encapsulation: for example, created or permalink properties. You really don't want to reimplement this all over again.

To link HPI package with your private data, you add a small section to your private config in ~/.config/my/my/config/__init__.py:

class roamresearch:
    export_path = '/exports/roamresearch/'
    username = 'covid19'

After this, you get IDE help when working with Roam data. For example, mypy will remind me that title is optional, so I won't have None related issues.

Interactive queries

At this stage, you can play with the data interactively in your Python interpreter. For example:

  • "what's the most active note in my database?"

    import my.roamresearch as RR
    roam = RR.roam()
    note = max(roam.notes, key=lambda n: len(n.children))
    print(note)
    
    Node(created=2020-02-27 02:29:22.725000+00:00, title=Worst-Case Scenario Planning, body=None)
    
  • "what's the most 'dangling' link in my database"

    This could be useful to organize it, and fill frequently occurring references with more context.

    The code is very straightforward:

    import my.roamresearch as RR
    roam = RR.roam()
    
    import re
    def all_internal_links():
        for note in roam.traverse(): # go recursively through all Roam outlines
            links = re.findall(r'\[\[(.*?)\]\]', note.body or '') # extract the [[links]]
            yield from links
    
    existing = { # figure out which links already have meaningful pages for them
        n.title for n in roam.traverse()
        if not n.empty()
    }
    missing = [
        link for link in all_internal_links()
        if link not in existing
    ]
    
    from collections import Counter
    print(Counter(missing).most_common(5))
    
    slider 22
    China 12
    Open Questions 10
    Travel 8
    COVID-19 6

    [[slider]] is a just a parsing artifact, but you can see that perhaps it's a good time to add some content on the "China" page!

This is cool because you don't have to wait till developers add new features or come up with a query language (which is also really hard!). You can use a powerful programming language you already know.

5 Use Orger for org-mode mirror

Now that accessing the data is a matter of import my.roamresearch, we can enjoy using it!

One very useful tool for interacting with my data is Orger. Orger is a library that helps you generate an org-mode representation of your data.

I won't go into the Orger module implementation details – you can check out the code here: roamresearch.py. There are also more modules for other services like Twitter, Reddit, Instapaper, etc.

We will use Orger to generate a read-only mirror of the Roam database. The idea is that you run Orger script periodically to get a reasonably up-to-date representation of your Roam data. You'll still use Roam web app to modify your notes.

Of course, it would be cool to have proper bidirectional sync, so your modifications to the org-mode file are reflected in Roam. But it's a much harder problem, especially without an API.

But even having a read-only replica enables you to do lots of cool things! I invite you to watch a short demo I recorded, and I'll highlight some features here in writing.

You benefit from the existing infrastructure and tooling around org-mode:

  • the org-mode markup and text editor integration

    Roam markup maps naturally into org-mode (perhaps not very surprising).

    So you can use outlines/colors/bold/italic/linka/code blocks/whatnot. As a bonus, you get all the associated Emacs customizations.

  • the Emacs navigation features

    You can use your favorite keybindings with no need to learn new hotkeys hardcoded by the developers. For example:

    • editor specific: with evil-mode you can just use Vim navigation (hjkl , $ / w / b / etc.)
    • org-mode specific: folding/unfolding (Shift+Tab/Tab), outline navigation ({ , }), jumping to the heading (gh)

      Again, if you're an evil person, just check out evil-org-mode.

    • Link navigation: you can use link-hint to quickly navigate links with your keyboard (SPC s l in Doom Emacs)

      This is similar to web browser extensions like vimperator, Tridactyl and Surfingkeys.

    Navigating in your text editor is very fast and you never experience network latency or backend hiccups.

  • if you need to edit the notes, you can jump back into the Roam web interface by following the permalink
  • [[Internal Links]] in Roam also have first-class support in org-mode, so you can jump within the file
  • the Emacs search modules

    • regular plaintext search (/ if you're a Vim person)
    • helm-swoop: shows you the search results and the context in which each result occurred.

      It's probably easier if you check out the demo gif than to put it in words.

    • helm-ripgrep: I'm using it for a global incremental search (a single keybinding away) over all of my personal data.

      And if you're still not using ripgrep, you should seriously give it a try.

    Typically, web apps provide a pretty pathetic search experience. Sometimes it's just not unavailable. When it is, it's slow, the results are paginated, ranked by some magic algorithm and the scope is often restricted for performance reasons. Roam actually stands out because unlike most, it does have an incremental search within the notes.

    Searching in Emacs is an incredibly pleasant experience. The search is incremental, updates results as fast as you type, and you can configure it as you wish.

    Warning: once you get used to a fast incremental search in Emacs, you will struggle to go back. You'll end up grumpy and as frustrated about web apps as me. Think twice.

And some extra things I haven't mentioned in the video:

  • you get (read-only) offline mode for free

    Roam does offer offline mode in the browser, but however talented Roam engineers are, a plain text file will always be more reliable.

    Fun fact: during the first attempt to record the demo, Roam web app wasn't loading :P

  • you get mobile access

    Roam doesn't have first-class mobile support yet, so with an org-mode mirror, you can benefit from the existing org-mode apps for your phone.

    Again, Roam specifically might implement it in the future, but the point is that you can be ahead of the developers and do it yourself.

if org-mode is so cool why not just use it instead of Roam?

Better ask someone who actively uses Roam, because I'm a stubborn org-mode user! If you're like me, you probably tend to stay away from web apps.

In that case, perhaps, check out jethrokuan/org-roam, which implements a workflow similar to Roam, but within Emacs.

But even I can come up with some objective (I hope) reasons to avoid Emacs:

  • entry barrier

    I won't lie, I spent a while grasping and learning Emacs and I still struggle at times.

    But also, most of the frustrations are when I'm trying to implement things that wouldn't even be possible in other apps and editors without modifying their source code or self-hosting. With a reasonable config, like doom emacs, you can get very far just on the defaults.

  • cloud sync

    As with all cloud apps, this a curse and a blessing at the same time.

    The "blessing" bit will be especially evident if Roam adds real-time collaboration (and I bet it's more "when" than "if").

    I have to admit, to my knowledge, org-mode gang doesn't have much to offer here. I think this is a very important problem and needs solving.

  • easier integration with web

    For example, you can embed tweets in Roam, and there is lots of potential for dynamic features. A browser frontend for Emacs might actually be a good idea.

Org-mode mirror can complement your Roam database and serve as an additional interface and a means of learning org-mode, without having to fully succumb to it.

6 Promnesia: integrate Roam with your browser

Promnesia is an extension for enhancing your browser history. It unifies the browsing history, which is siloed in various apps and services; and lets you explore the history in context, showing the information on where and how you've encountered the links. If you want to learn more about it, check out demos, here I'll only concentrate on demonstrating Roam.

To use Roam data in Promnesia, all you need is to add a small bit to the config:

from promnesia.sources import roamresearch
SOURCES = [
   ...,
   Source(roamresearch.index, name='roam'),
   ...,
]

And that's it! This enables Roam Research module (40 lines of code), which consumes the data from HPI.

You can also index others' public Roam databases. If Roam is a second brain, this can be a way of interfacing with others' second brains!

I've recorded a demo, showing various workflows for your Roam data, which are enabled by Promnesia.

Here's a text summary of the features, demonstrated in the video:

  • the sidebar

    Shows your browsing history for the current page along with the source, timestamp, and the context, in which the link occurred.

    Clicking on the permalink brings us straight to the outline within the Roam Research web app.

  • what are the github repositories saved in this Roam database?

    If you open the sidebar on github.com, you can see the subpages along with the contexts, in particular, links to all the github repositories saved in Roam.

  • what are the tweets saved in this Roam database

    Similarly, if you open twitter.com, you can see all the tweets in your browsing history. If you open a specific Twitter profile, you get to see all the tweets from a specific person.

    I find it extremely useful to recall when you've engaged with the content from a specific person, for example, to figure out whether you should follow them.

  • full text search over your browsing history
  • how did I get on this page?

    You can trace through all of your history to remember why you bookmarked a page.

  • what links on this page have I already visited?

    When you press 'mark visited', Promnesia highlights the links you've already visited so you can explore new information more efficiently.

There is some room for improvement, especially in terms of the UI (e.g. markup handling could be better), but these are easy to fix.

These workflows aren't unique to Roam, and possible with many other data sources. Promnesia already supports many, and adding new is very easy. You can find find more links, screenshots, and demos in the repository.

Thanks for watching, check it out, and let me know if you can think of other cool features!

7 --

Happy to hear your feedback or help you set up the tools I mention!