diff --git a/assets/images/maghammer.png.original b/assets/images/maghammer.png.original
index 43c3cbc..c1920e1 100644
Binary files a/assets/images/maghammer.png.original and b/assets/images/maghammer.png.original differ
diff --git a/assets/images/maghammer2.png.original b/assets/images/maghammer2.png.original
new file mode 100644
index 0000000..c0e6276
Binary files /dev/null and b/assets/images/maghammer2.png.original differ
diff --git a/assets/images/maghammer_v2_record.png b/assets/images/maghammer_v2_record.png
new file mode 100644
index 0000000..11e01d5
Binary files /dev/null and b/assets/images/maghammer_v2_record.png differ
diff --git a/assets/images/maghammer_v2_search.png b/assets/images/maghammer_v2_search.png
new file mode 100644
index 0000000..1b3d612
Binary files /dev/null and b/assets/images/maghammer_v2_search.png differ
diff --git a/assets/images/maghammer_v2_table.png b/assets/images/maghammer_v2_table.png
new file mode 100644
index 0000000..b0a7efe
Binary files /dev/null and b/assets/images/maghammer_v2_table.png differ
diff --git a/assets/images/tokenization.png b/assets/images/tokenization.png
new file mode 100644
index 0000000..22b3456
Binary files /dev/null and b/assets/images/tokenization.png differ
diff --git a/blog/maghammer.md b/blog/maghammer.md
index da93438..b76f930 100644
--- a/blog/maghammer.md
+++ b/blog/maghammer.md
@@ -1,9 +1,11 @@
---
title: "Maghammer: My personal data warehouse"
created: 28/08/2023
-updated: 12/09/2023
+updated: 14/05/2024
description: Powerful search tools as externalized cognition, and how mine work.
slug: maghammer
+series: maghammer
+series_index: 1
---
::: epigraph attribution="Deus Ex"
The need to be observed and understood was once satisfied by God. Now we can implement the same functionality with data-mining algorithms.
@@ -49,6 +51,7 @@ Currently, I have custom scripts to import this data, which are run nightly as a
* [Miniflux](/rssgood/) RSS feed entries.
* [Minoteaur](/minoteaur/) notes, files and structured data. I don't have links indexed since SQLite isn't much of a graph database[^6], and my importer reads directly off the Minoteaur database and writing a Markdown parser would have been annoying.
* RCLWE web history (including the `circache` holding indexed pages in my former Recoll install).
+* Emails dumped from Thunderbird mailboxes (I really did not enjoy writing the parser for that format).
There are also some other datasets handled differently, because the tools I use for those happened to already use SQLite somewhere and had reasonably usable formats. Specifically, [Gadgetbridge](https://www.gadgetbridge.org/) data from my smartwatch is copied off my phone and accessible in Datasette, [Atuin](https://github.com/ellie/atuin)'s local shell history database is symlinked in, Firefox history comes from [my script](https://github.com/osmarks/random-stuff/blob/master/histretention.py) on my laptop rather than the nightly serverside batch job, and I also connected my Calibre library database, though I don't actually use that. 13GB of storage is used in total.
@@ -81,7 +84,7 @@ It is actually somewhat more complex than that for various reasons. I had to mod
## Future directions
-The system is obviously not perfect. As well as some minor gaps (browser history isn't actually put in a full-text table, for instance, due to technical limitations), many data sources (often ones with a lot of important content!) aren't covered, such as my emails and conversation history on e.g. Discord. I also want to make better use of ML - for instance, integrating things like Meme Search Engine better, ~~local Whisper autotranscription of videos rather than having no subtitles or relying on awful YouTube ones, semantic search to augment the default [SQLite FTS](https://www.sqlite.org/fts5.html) (which uses term-based ranking - specifically, BM25),~~ and OCR of screenshots. I still haven't found local/open-source OCR which is both good, generalizable and usable[^3]. Some of the trendier, newer projects in this space use LLMs to do retrieval-augmented generation, but I don't think this is a promising direction right now - available models are either too dumb or too slow/intensive, even on GPU compute, and in any case prone to hallucination.
+The system is obviously not perfect. As well as some minor gaps (browser history isn't actually put in a full-text table, for instance, due to technical limitations), many data sources (often ones with a lot of important content!) aren't covered, such as conversation history on e.g. Discord. I also want to make better use of ML - for instance, integrating things like Meme Search Engine better, ~~local Whisper autotranscription of videos rather than having no subtitles or relying on awful YouTube ones, semantic search to augment the default [SQLite FTS](https://www.sqlite.org/fts5.html) (which uses term-based ranking - specifically, BM25),~~ and OCR of screenshots. I still haven't found local/open-source OCR which is both good, generalizable and usable[^3]. Some of the trendier, newer projects in this space use LLMs to do retrieval-augmented generation, but I don't think this is a promising direction right now - available models are either too dumb or too slow/intensive, even on GPU compute, and in any case prone to hallucination. After some time in use use, it seems like the most critical thing to change is how chunks for embedding are generated and organized: a chunk from midway through a document retains no context about the title or other metadata, the ranker doesn't aggregate multiple chunks from within the document properly, and in my laziness (not wanting to bring in a tokenizer) they're way shorter than they have to be.
Another interesting possibility for a redesign I have is a timeline mode. Since my integration plugin (mostly) knows what columns are timestamps, I could plausibly have a page display all relevant logs from a day and present them neatly.
diff --git a/blog/maghammer2.md b/blog/maghammer2.md
new file mode 100644
index 0000000..738c22e
--- /dev/null
+++ b/blog/maghammer2.md
@@ -0,0 +1,81 @@
+---
+title: Lessons learned from Maghammer
+created: 06/07/2024
+description: I got annoyed and rewrote everything.
+slug: maghammer2
+series: maghammer
+series_index: 2
+---
+::: epigraph attribution=RossM
+If Google Search does not return the result the user wanted because the user typed the wrong thing in the search box, that is a failure on Google's part, not a failure on the user's part. Conversely, if Google Search does return the result the user wanted because the user typed the correct thing in the search box, that is a failure on Google's part, not a failure on the user's part.
+:::
+
+[Maghammer](/maghammer/), the search and data management system I previously outlined, has been reasonably successful, as it's frequently been able to find things which were hard to look up any other way. But, like all software, it has problems. It has somewhat more than usual, being an assemblage of very ugly Python scripts made out of expediency and laziness. Many of these were knowable in advance if I had paid more attention, but I have at least gained some interesting information from its slightly-less-than-a-year of production use. Due to vague dissatisfaction with Python and mounting implementation problems, I've done a full rewrite in Rust. As the code has now gone from "active affront to God" to merely "quite inelegant", I will be releasing it shortly (though I do not expect it to be very usable for other people). Here's what I've determined and changed:
+
+## Postgres and PGVector are very good
+
+::: epigraph attribution=ryg
+Give someone state and they'll have a bug one day, but teach them how to represent state in two separate locations that have to be kept in sync and they'll have bugs for a lifetime.
+:::
+
+RDBMSes like PostgreSQL and SQLite are pieces of software I very much appreciate. Their data model may not nicely fit everything I do, but in an industry where many chase after convoluted solutions to problems they do not actually have they are a bright point - they Just Work, centralizing data such that I can access it across languages and manually using pleasant tools, and have excellent consistency guarantees. When it's possible to use them as the only location my code persists state, I will generally do that for the sheer simplicity and reliability gains.
+
+Unfortunately, the old version of Maghammer was *not* built like this. Its search index was built on [SQLite FTS5](https://www.sqlite.org/fts5.html), which for some bizarre reason requires manually[^1] synchronizing main table content with FTS table content on pain of integrity failures, using an oddly procedural API, and with a separate FAISS index for the semantic search component which also had to be periodically kept in sync. This led to a lot of annoying code and some minor but annoying bugs.
+
+Postgres has no competent lexical (word-based) full text search built in[^2], and I haven't yet found an extension providing this in a satisfying way, but it does have [PGVector](https://github.com/pgvector/pgvector), which can store and index vectors. I had previously evaluated it for Maghammer, but determined that it was unsuitable - ANN indices were problematically slow to build, and without half-precision vector support it used lots of storage and RAM unnecessarily.
+
+However, with PGVector getting new features (half-precision vector support and faster but more storage-hungry HNSW indices), me realizing that I had not properly performance-tuned Postgres[^3], and some fixes to the embedding generation process itself, it became practical to use it as the main vector store, so my rewritten version stores all its data in Postgres (rather than a variety of SQLite databases and a FAISS index) along with the embeddings.
+
+## Smarter tokenization and embeddings
+
+One significant problem with the last version is that the context window used for embeddings was short, so queries which would only match given a large window of the text returned irrelevant results. This is because tokenization is terrible[^4] and I avoided having to deal with it in the indexing code by chunking text by sentence until a character length limit was exceeded. In all reasonable circumstances (I lose no sleep over discarding weird Unicode) this kept it under the length limit, but usually very far under the length limit. The new code is tokenization-aware and breaks inputs into chunks by tokens, which results in oddly placed breaks between chunks but makes them much longer.
+
+This has the additional advantage of producing fewer sentence embedding vectors, which makes the index more efficient.
+
+As a minor consequence of picking from models released about a year later, I also replaced the [e5-large-v2](https://huggingface.co/intfloat/e5-large-v2) model previously used for embeddings with the same-sized-and-apparently-better [snowflake-arctic-embed-l](https://huggingface.co/Snowflake/snowflake-arctic-embed-l), since it claims better retrieval performance and is more trustworthy, to me, than the Chinese models also claiming that.
+
+Both models use a prefix to indicate whether an input is a query or a passage to match against, but the newer one seems to be more sensitive to them (or it could simply be the longer inputs), so I've also split columns into "short" and "long" to determine whether this prefixing mechanism is used for queries or not - without this, short passages are privileged, especially ones containing, for some ridiculous reason[^5], the literal text `passage`. This has its own problems, so I might need an alternative solution.
+
+## The quantitative data is not all that helpful
+
+While I could in principle get interesting results out of analyzing things like web browsing activity by day and whether my step count is correlated with anything interesting, I have not had any compelling reason to do this yet, and this would likely require complex dedicated analysis scripts and frontends to do well. Datasette's ability to run and nicely render custom SQL queries is cool, but not very relevant to this - I've only ever used it about five times in total.
+
+## Sharded vector indices
+
+Previously, all embeddings from all columns of all tables indexed by the system were stored in a single FAISS index and then resolved to their original sources after a query was done. I'm not actually sure why I did this - presumably because it made the embedding generation code simpler - but this had a number of problematic consequences.
+
+None of the ANN index algorithms in wide use[^6] are able to efficiently handle simultaneously searching by the vector component and normal RDBMS-like columns - if you want to find all good matches with a certain tag, or after a certain date, for instance, you have to do a linear scan through the rows returned from your vector search, which degrades to a linear scan through your entire database as satisfying results become rarer. The upshot, for my application, is that it was not possible to efficiently filter search results by which type of data (media subtitles, personal notes, archived PDFs, etc) they came from.
+
+By instead creating separate tables and thus indices for each table and each column of text, and querying them in parallel, it's possible to cheaply get results from specific text columns, though of course this does not generalize to queries splitting by other fields.
+
+## Most of Datasette is unnecessary, at least for me for this
+
+The choice to use Datasette for the UI and scaffolding of the original version was deliberately made to save on development time and build a mostly working prototype quickly. It would have been impractical for me to reimplement all of it on hacky Rust, but I didn't actually have to: the features I actually used most were general text search, LIKE queries on some columns, viewing individual records and very basic table filtering. This was much easier to cover than the entire featureset.
+
+## Making a special-purpose tool gives me more flexibility
+
+Datasette has to be able to usefully read off arbitrary SQLite databases regardless of their structure, and SQLite databases don't encode that much semantic information (there's not even a dedicated type for timestamps, and I've seen everything from RFC 3399 strings to nanoseconds since the Unix epoch to the [Julian day](https://www.sqlite.org/lang_datefunc.html)). In the rewrite, all tables are explicitly populated by import logic in the program and all tables have some metadata attached so the code can know which columns should be shown in a tabular table view and which contain meaningfully searchable text. This will make it easier to add features like the previously-alluded-to timeline mode. Perhaps it will also be practical to add nice graphs later.
+
+I also got to redesign the UI, inasmuch as I wrote all the UI code and markup from scratch. It's based on the "grey borders and rectangles" design language used for most of my internal tools.
+
+::: captioned src=/assets/images/maghammer_v2_table.png
+The table view showing some shell history.
+:::
+::: captioned src=/assets/images/maghammer_v2_record.png
+An example record view: unlike in Datasette, it is laid out vertically for longer columns.
+:::
+::: captioned src=/assets/images/maghammer_v2_search.png
+Search results even get colors (by indexer) and will link to the source where available (in this case, the local copy of the video) and the available record.
+:::
+
+[^1]: Or using triggers. I did not do it this way since there had to be non-SQL change tracker code anyway for the semantic search component.
+
+[^2]: The built-in `tsvector`/`tsquery` system lacks good ranking and has performance issues. SQLite's is weird to use, but very good at its job.
+
+[^3]: By default, it seems to assume you have a computer from the 1990s, and I hadn't fully corrected this.
+
+[^4]: 
+
+[^5]: I think this is because the overly long instruction for queries contains "passage".
+
+[^6]: There do exist ways to fix this but they're either bad or tricky mathematics.
\ No newline at end of file
diff --git a/package-lock.json b/package-lock.json
index 2b5c783..346be38 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -11,7 +11,7 @@
"dependencies": {
"@msgpack/msgpack": "^3.0.0-beta2",
"axios": "^1.5.0",
- "better-sqlite3": "^8.7.0",
+ "better-sqlite3": "^11.0.0",
"chalk": "^4.1.0",
"dayjs": "^1.8.28",
"esbuild": "^0.19.6",
@@ -526,10 +526,11 @@
]
},
"node_modules/better-sqlite3": {
- "version": "8.7.0",
- "resolved": "https://registry.npmjs.org/better-sqlite3/-/better-sqlite3-8.7.0.tgz",
- "integrity": "sha512-99jZU4le+f3G6aIl6PmmV0cxUIWqKieHxsiF7G34CVFiE+/UabpYqkU0NJIkY/96mQKikHeBjtR27vFfs5JpEw==",
+ "version": "11.0.0",
+ "resolved": "https://registry.npmjs.org/better-sqlite3/-/better-sqlite3-11.0.0.tgz",
+ "integrity": "sha512-1NnNhmT3EZTsKtofJlMox1jkMxdedILury74PwUbQBjWgo4tL4kf7uTAjU55mgQwjdzqakSTjkf+E1imrFwjnA==",
"hasInstallScript": true,
+ "license": "MIT",
"dependencies": {
"bindings": "^1.5.0",
"prebuild-install": "^7.1.1"
@@ -2251,9 +2252,9 @@
"integrity": "sha512-AKpaYlHn8t4SVbOHCy+b5+KKgvR4vrsD8vbvrbiQJps7fKDTkjkDry6ji0rUJjC0kzbNePLwzxq8iypo41qeWA=="
},
"better-sqlite3": {
- "version": "8.7.0",
- "resolved": "https://registry.npmjs.org/better-sqlite3/-/better-sqlite3-8.7.0.tgz",
- "integrity": "sha512-99jZU4le+f3G6aIl6PmmV0cxUIWqKieHxsiF7G34CVFiE+/UabpYqkU0NJIkY/96mQKikHeBjtR27vFfs5JpEw==",
+ "version": "11.0.0",
+ "resolved": "https://registry.npmjs.org/better-sqlite3/-/better-sqlite3-11.0.0.tgz",
+ "integrity": "sha512-1NnNhmT3EZTsKtofJlMox1jkMxdedILury74PwUbQBjWgo4tL4kf7uTAjU55mgQwjdzqakSTjkf+E1imrFwjnA==",
"requires": {
"bindings": "^1.5.0",
"prebuild-install": "^7.1.1"
diff --git a/package.json b/package.json
index d5978b7..afc9714 100644
--- a/package.json
+++ b/package.json
@@ -6,7 +6,7 @@
"dependencies": {
"@msgpack/msgpack": "^3.0.0-beta2",
"axios": "^1.5.0",
- "better-sqlite3": "^8.7.0",
+ "better-sqlite3": "^11.0.0",
"chalk": "^4.1.0",
"dayjs": "^1.8.28",
"esbuild": "^0.19.6",
diff --git a/src/global.json b/src/global.json
index 4344ca5..3703c06 100644
--- a/src/global.json
+++ b/src/global.json
@@ -23,7 +23,8 @@
"I am a transhumanist because I do not have enough hubris not to try to kill God.",
"If at first you don't succeed, destroy all evidence that you tried.",
"One man’s constant is another man’s variable.",
- "All processes that are stable we shall predict. All processes that are unstable we shall control."
+ "All processes that are stable we shall predict. All processes that are unstable we shall control.",
+ "This website doesn't exist. This is a Boltzmann website."
],
"feeds": [
"https://www.science.org/blogs/pipeline/feed",
diff --git a/src/index.js b/src/index.js
index 0ecb139..85aced3 100644
--- a/src/index.js
+++ b/src/index.js
@@ -239,18 +239,44 @@ const processExperiments = async () => {
}
const processBlog = async () => {
- const templates = globalData.templates
const blog = await loadDir(blogDir, async (file, basename) => {
- return applyTemplate(templates.blogPost, file, async page => {
- const out = path.join(outDir, page.data.slug)
- await fse.ensureDir(out)
- return path.join(out, "index.html")
- }, { processMeta: (meta, page) => {
- meta.slug = meta.slug || removeExtension(basename)
- meta.wordCount = page.content.split(/\s+/).map(x => x.trim()).filter(x => x).length
- meta.haveSidenotes = true
- }, processContent: renderMarkdown })
+ const page = parseFrontMatter(await readFile(file))
+ const meta = page.data
+ meta.slug = meta.slug || removeExtension(basename)
+ meta.wordCount = page.content.split(/\s+/).map(x => x.trim()).filter(x => x).length
+ meta.haveSidenotes = true
+ meta.content = renderMarkdown(page.content)
+ return meta
})
+
+ const series = {}
+ for (const [name, data] of Object.entries(blog)) {
+ if (data.series_index) {
+ series[data.series] ??= []
+ series[data.series].push({ index: data.series_index, post: name })
+ }
+ }
+ for (const entries of Object.values(series)) {
+ entries.sort((a, b) => a.index - b.index)
+ for (let i = 0; i < entries.length; i++) {
+ const currentEntry = blog[entries[i].post]
+ if (i > 0) {
+ currentEntry.prev = blog[entries[i - 1].post]
+ }
+ if (i + 1 < entries.length) {
+ currentEntry.next = blog[entries[i + 1].post]
+ }
+ }
+ }
+ for (const page of Object.values(blog)) {
+ const out = path.join(outDir, page.slug)
+ await fse.ensureDir(out)
+ await fsp.writeFile(path.join(out, "index.html"), globalData.templates.blogPost({
+ ...globalData,
+ ...page
+ }))
+ }
+
console.log(chalk.yellow(`${Object.keys(blog).length} blog entries`))
globalData.blog = addGuids(R.filter(x => !x.draft && !x.internal, R.sortBy(x => x.updated ? -x.updated.valueOf() : 0, R.values(blog))))
}
diff --git a/src/style.sass b/src/style.sass
index 2aa5eb5..61561c5 100644
--- a/src/style.sass
+++ b/src/style.sass
@@ -305,6 +305,9 @@ $hl-border: 3px
max-width: 15em
display: block
+.sidenotes img.big
+ max-width: 30em
+
.hoverdefn
text-decoration-style: dotted
text-decoration-line: underline
@@ -352,4 +355,7 @@ table
// TODO
#comments-wrapper textarea
- width: calc(100% - 0.5em) !important
\ No newline at end of file
+ width: calc(100% - 0.5em) !important
+
+.next
+ text-align: right
\ No newline at end of file
diff --git a/templates/layout.pug b/templates/layout.pug
index 415184a..65fd98e 100644
--- a/templates/layout.pug
+++ b/templates/layout.pug
@@ -62,6 +62,16 @@ html(lang="en")
span= `${metricPrefix(wordCount, "")} words`
if description
em.description!= description
+ if prev
+ div
+ .prev
+ span= "← "
+ a(href=`/${prev.slug}`)= prev.title
+ if next
+ div
+ .next
+ a(href=`/${next.slug}`)= next.title
+ span= " →"
block content
if haveSidenotes