I think what we're seeing with every database company providing new full-text search capabilities is an example of AI coding productivity showing up in the real world.
AFAIK all of these are implementations of the BM25 algorithm. You can just tell an agent to read about BM25 and implement it in your system of choice. Cool to see. Seems like there's still a lot of juice to be squeezed out of how it's architected and integrated into each system, but you can't help but wonder if this will lead to aggressive commodification
Was not able to find any mention of AI or LLM usage on the article. The article is very detailed and goes in depth about how they have been able to do it. If anything it just shows the database level expertise and understanding of the existing implementations to find the optimization opportunities.
Unless its explicitly mentioned lets not dilute the credit of the folks who worked on.
There is a lot of truth to this, but it's also very much down to domain experts being able to do this to move faster.
Planetscale (assuming they used a agentic development practice) will have pulled this off, to the level of performance that they have, because they have a team of very highly experienced Postgres developers. Their knowlage of Postgres internals will have given them the insights needed to steer the models to a plan that used the architecture as described in the post. That's not something a model can do on its own*
World experts + LLMs = moving mountains.
(* we're obviously seeing something a little different from inside the research teams in the labs. They are showing that the models, when you burn the level of tokens only they can, are able to do novel things from the models own insights.)
I am exploring this exact area of search and analytics for vanilla postgres as replicas. Guess what, the LLM came up with this exact conclusion of using ctids as docids, all by itself. It was surreal for me to read the blog above , when I hit that paragraph about ctids.
> There is a lot of truth to this, but it's also very much down to domain experts being able to do this to move faster.
Yeah, I don't think I could tell Qwen3.8 (my LLM of choice) to study up on bm25 and then implement full text search in the couchdb instances I maintain without studying both bm25 and couchdb internals myself.
Seems like a lot of this knowledge was encoded into the blog post. I wonder if given this post and access to a planet scale instance to compare with, how close an agentic agent could get.
I think we can frame it as LLMs materializing existing potential. It seems like there needs to be an underlying potential to tap into, without which, the results could be slop.
I don't completely disagree with your hypotheses but it feels like the hard part of his TIN stuff isn't BM25 (which has been around for donkeys years) it's all the hardcore storage engine work around it. And is an LLM particularly good at e.g. segment merging under a thousand updates a second? I've had a few situations where I've been told "we've hit the perf floor" by Claude only to have persisted myself and shaved substantial amounts off still.
More damning for the theory might be that I think paradedb's pg_search predates the agentic coding by a few years?
Becoming more the norm for them, Neki is the same.
Immediately rules out ever using them (though I don't currently have any problems that would benefit from that level of scale currently, have in the past though).
Postgres's license allows this but for me (personally) it leaves a bad taste.
Also it's not really "full-text search for Postgres" it's "full-text search for our hosted version of Postgres" so the title is a little misleading.
> Becoming more the norm for them, Neki is the same.
Uh, isn’t this becoming the norm everywhere ever since LLMs have been trained on OSS without credit or attribution? Why wouldn’t you want to hide your stuff going forward?
In my view, OSS is only going to move more and more towards one of two models: open core + proprietary functionality (e.g. MongoDB) OR open source + private tests (e.g. SQLite).
Building non-open extensions on top of it, it's not the license I don't like, the bad taste is that they use something open extend it and keep part of it closed.
The license allows it but on the flip side it's vendor lock-in predicated on using something open as the base.
Fully proprietary no issue with that, full open, no issue with that, building proprietary on top of open is where the bad taste comes in.
For completeness, it's not them specifically either, the other cloud companies do similar things and I suspect in part the reason they don't open these extensions up is because the others will but then they are doing the same thing themselves.
I find this attitude self-defeating because the lure of being able to provide some amount of proprietary software on top of open is what draws in the corporate investment in open software. And in the system we live in, it’s hard to imagine there’d be nearly as much open software as there is without that corporate investment. As a big fan of open software, this seems like a great trade to me.
(Disclosure: I work for a company with this business model, in part because I like working on open software)
When one develops using open-source software they have an obligation to follow the licenses. They also have a moral obligation to be respectful of the work upon which they’re building. And they have a social obligation to help improve that software where they can.
Those that develop on top of open-source have no obligation to give you their work for free.
You’d be surprised as to the amount of open-source contributions TIN drove towards Postgres, LLVM, and pgrx. And you’d be speechless at the amount of upstream work across all sorts of open-source PlanetScale does. Postgres 18.6, for example, is better for you today, in part, because of TIN. You’re welcome.
I didn't say thank you and don't presume I would, Planetscale acting in their own self interest by improving postgres upstream isn't deserving of thanks, any more than Intel upstreaming a bunch of Linux kernel work is or myriad other examples.
Corporations acting in their self interest isn't worth giving thanks for, neither is the work of the people paid to do work on their behalf.
I don't expect my employer to thank me, I expect them to pay me, I don't expect users of software I was paid to write to thank me because I did it for the money not out of altruism towards those users.
> Those that develop on top of open-source have no obligation to give you their work for free.
This is a pretty narrow view of open source.
Just as a simple example, GPLv3 and AGPLv3 are both considered Open Source and, depending on how you hold them, may obligate releasing work to customers essentially for free.
"...are both considered Open Source" << Free Software (GPL) is the O.G. - it's like saying "George Washington is considered to be an American President" or "The Beatles are considered to be a pop/rock band" or "water is considered to be..."
I remember so well when "Open Source" branding started with Bruce Perens - all the business arguments. When you need a database, someone ELSE'S business decisions (how THEY are going to make money) never end up helping YOU. If you use proprietary extensions and become dependent on them, you inevitably will get BURNED when their business needs diverge from your needs.
- They close shop
- They refuse to interop with something you need
- They demand that you obey their arcane rules
- They rug-pull
- They get hacked as only they can
- They lie to you
- They stab you in the back
> When one develops using open-source software they have an obligation to follow the licenses.
Agreed!
> They also have a moral obligation to be respectful of the work upon which they’re building.
Respectfully, completely disagree!
Open source is a license, not a moral framework. Your only duty is to abide by the license. If you want to enforce that everything built on top of a particular open source package also be open source, that belongs in the license.
We live in a society. I don’t see any harm in believing we ought to be respectful of the work. That’s all I’m saying. You don’t have to be, but it’s a better world if you are.
> If you want to enforce that everything built on top of a particular open source package also be open source, that belongs in the license.
Yes! I fully agree. The one that creates the thing is the one that gets to choose its license.
Having opened-sourced some bit of work myself, that’s a very difficult decision.
Postgres does have pg_fts (tsvector/tsquery/tsrank) which is a quite sophisticated full text search package integrated with functional indexing and query optimization. Why would I use something vibecoded that isn't part of core Postgres instead?
PostgreSQL built-in FTS provides a score for each row based just on the data for that row.
Relevance algorithms like BM25 take overall corpus statistics into account. If you search for a bunch of words and some of them are less common than others in the overall set of documents, documents that match THOSE words will score higher than matches for other words in your search.
That's what all of these additional extensions are providing.
If you read deep into this, they claim much better performance than the built in search; they also imply that the built-in search is missing features they provide but don’t make clear which ones (I think it is just support in the same index for queries covering other conditions on other columns, because every other feature they claim seems to line up with the built in search features, which have been around for about 20 years.)
- superior performance
- superior operational overhead
- no second copy of data in tsvector form
- BM25 scoring support with optimized top-k output
- runtime configurable scoring knobs
- expression-attached score boosting
- sophisticated span query support -- this is proximity search on steroids (https://github.com/planetscale/lead/tree/main/tinql/docs)
- lossless term positions
- index-answerable negative expressions (find all docs that don't contain a word)
- full document hit highlighting
- optimized exact `count(\*)`
- term expansion via any of fuzzy matching, wildcards, regular expressions, and dictionary ranges
- intentionally smaller user-facing SQL API surface
There's a lot we didn't cover in the announcement blog. I'm sure we'll do more as time goes on.
As an aside, something I personally think is cool, and I suppose you can do this with Postgres' built-in `@@` too, is that you can use TIN's full query language (linked above) against any text datum. This is a valid query:
SELECT pid, query
FROM pg_stat_activity
WHERE query ==> 'select OR copy'
in other words, you don't need an index at all to use TIN's full query language against any text field in any query.
The built in search can't do any scoring mechanism that involves corpus-wide stats, so things like tfidf and bm25 are right out. If you don't need that then great, but in my experience the results are much worse.
Interestingly enough SQLites FTS supports Lucene queries out of the box with great performance characteristics. IIRC only writes become pretty slow after a while.
I’ve always wondered what exactly would prevent PostgreSQL from strapping that implementation into its own database. My experience with ts_query hasn’t been particularly rosy. It can be better than LIKE but only marginally so and at the cost of insane index sizes…
If this extension becomes open source and we can test it out in the real world I’m sure there’s a sweet spot
SQLite FTS relies on shadow B-trees under single-writer locks. Postgres index access methods must map postings directly to physical ctid tuples, surviving MVCC visibility checks and heap tuple churn.
There's another aspect which none of the FTS search solutions for Postgres do well in my opinion: multi-language support.
For example this one: it doesn't mention support for CJK languages (meaning tokenization for e.g. Chinese will resolve to one token per character, which will technically work and give results, but is inefficient). Also word stemming (databases -> database) is also missing as far as I can see, so the kind of queries where you'd expect related words to show up will be missing. Just doing case-folding and accent-folding is a bit of a functional but bruteforce solution.
Ideally I'd want something that supports:
- language aware tokenization, with ability to define the language per record. Including stemming, etc. And have useful predefined configuration for common languages (e.g. the Postgres built in one is missing many languages).
- CJK support, tokenizing at word boundaries.
- Optional accent- and case-folding.
Most solutions just seem to assume English content, I have not found anything that does all of this yet.
I struggle with FTS inside SQL (SQLite and MSSQL). There is often a fairly significant impedance mismatch between the relational concerns and how the documents need to be stored.
I've always preferred to use SQL as the system of record and then build/maintain an external Lucene index. Do we think these integral FTS capabilities are at the point where a hybrid architecture doesn't make sense anymore? How much customization exists in this provider?
I’m one of TIN’s developers and if you google my username you’ll see I’ve been in this space for a long time.
The answer to your first question is simply: yes
As far as your second question, what customization do you need that you believe TIN or PlanetScale doesn’t provide? These are things we can do, with alacrity.
in my experience it’s pretty common to find big inverted indexes for text directly in the database - not necessarily large docs but certainly free text records in volume. using bm25 and unicode’s breakiterator is a very good way to build it. like putting lucene in the database basically - makes a lot of sense when the database is already large. places that bend over backwards to move search out of the db are usually trying to avoid having a very large db (and often end up with one anyway, getting the worst of both worlds)
They also end up with all the infrastructure and processes necessary to keep the external search system in sync, resync/reindex, pkey shipping back to their source of truth in queries, application-side joins and enrichment between both sources. It’s brutal.
Having everything in one place eliminates entire classes of development and especially operational problems.
I suppose it all depends on the scale of your project, but I've had pretty good luck using both MySQL's and SQLite's FTS capabilities. Surprised to hear that Open Source champion Postgres didn't have up-to-snuff FTS up to now...?
From the benchmarks deep in the document, TIN is much faster than built in text search (tested against GIN, which is itself much faster than GiST for text search.)
I think what we're seeing with every database company providing new full-text search capabilities is an example of AI coding productivity showing up in the real world.
It started with paradeDB and pg_search https://www.paradedb.com/blog/introducing-search
Timescale has pg_textsearch https://github.com/timescale/pg_textsearch
Neon and Databricks have Lakebase Search https://docs.databricks.com/aws/en/oltp/projects/lakebase-se...
Now PlanetScale.
AFAIK all of these are implementations of the BM25 algorithm. You can just tell an agent to read about BM25 and implement it in your system of choice. Cool to see. Seems like there's still a lot of juice to be squeezed out of how it's architected and integrated into each system, but you can't help but wonder if this will lead to aggressive commodification
Was not able to find any mention of AI or LLM usage on the article. The article is very detailed and goes in depth about how they have been able to do it. If anything it just shows the database level expertise and understanding of the existing implementations to find the optimization opportunities.
Unless its explicitly mentioned lets not dilute the credit of the folks who worked on.
There is a lot of truth to this, but it's also very much down to domain experts being able to do this to move faster.
Planetscale (assuming they used a agentic development practice) will have pulled this off, to the level of performance that they have, because they have a team of very highly experienced Postgres developers. Their knowlage of Postgres internals will have given them the insights needed to steer the models to a plan that used the architecture as described in the post. That's not something a model can do on its own*
World experts + LLMs = moving mountains.
(* we're obviously seeing something a little different from inside the research teams in the labs. They are showing that the models, when you burn the level of tokens only they can, are able to do novel things from the models own insights.)
LLMs are becoming a world expert in everything.
I am exploring this exact area of search and analytics for vanilla postgres as replicas. Guess what, the LLM came up with this exact conclusion of using ctids as docids, all by itself. It was surreal for me to read the blog above , when I hit that paragraph about ctids.
I am no postgres internals expert.
> There is a lot of truth to this, but it's also very much down to domain experts being able to do this to move faster.
Yeah, I don't think I could tell Qwen3.8 (my LLM of choice) to study up on bm25 and then implement full text search in the couchdb instances I maintain without studying both bm25 and couchdb internals myself.
Oh, you could.
Seems like a lot of this knowledge was encoded into the blog post. I wonder if given this post and access to a planet scale instance to compare with, how close an agentic agent could get.
The easiest way to find that out is to TIAS
THIS++
I think we can frame it as LLMs materializing existing potential. It seems like there needs to be an underlying potential to tap into, without which, the results could be slop.
ParadeDB's implementation builds on the Tantivy crate, which predates AI coding.
I don't completely disagree with your hypotheses but it feels like the hard part of his TIN stuff isn't BM25 (which has been around for donkeys years) it's all the hardcore storage engine work around it. And is an LLM particularly good at e.g. segment merging under a thousand updates a second? I've had a few situations where I've been told "we've hit the perf floor" by Claude only to have persisted myself and shaved substantial amounts off still.
More damning for the theory might be that I think paradedb's pg_search predates the agentic coding by a few years?
If anyone's curious - https://planetscale.com/docs/postgres/search/get-started#loc...:
They're not providing a local extension with the same performance at the time - it's only offered on their cloud services.
The local version https://github.com/planetscale/lead is mainly just for testing the syntax, it doesn't have the same perf characteristics.
Becoming more the norm for them, Neki is the same.
Immediately rules out ever using them (though I don't currently have any problems that would benefit from that level of scale currently, have in the past though).
Postgres's license allows this but for me (personally) it leaves a bad taste.
Also it's not really "full-text search for Postgres" it's "full-text search for our hosted version of Postgres" so the title is a little misleading.
> Becoming more the norm for them, Neki is the same.
Uh, isn’t this becoming the norm everywhere ever since LLMs have been trained on OSS without credit or attribution? Why wouldn’t you want to hide your stuff going forward?
In my view, OSS is only going to move more and more towards one of two models: open core + proprietary functionality (e.g. MongoDB) OR open source + private tests (e.g. SQLite).
why?
Why would you need super fast search for local testing?
To avoid vendor lock-in.
It's Postgres search any coding agent can switch you to something else in 5 minutes
Why don’t you like Postgres’ license? It’s as permissive as a license gets.
Building non-open extensions on top of it, it's not the license I don't like, the bad taste is that they use something open extend it and keep part of it closed.
The license allows it but on the flip side it's vendor lock-in predicated on using something open as the base.
Fully proprietary no issue with that, full open, no issue with that, building proprietary on top of open is where the bad taste comes in.
For completeness, it's not them specifically either, the other cloud companies do similar things and I suspect in part the reason they don't open these extensions up is because the others will but then they are doing the same thing themselves.
I find this attitude self-defeating because the lure of being able to provide some amount of proprietary software on top of open is what draws in the corporate investment in open software. And in the system we live in, it’s hard to imagine there’d be nearly as much open software as there is without that corporate investment. As a big fan of open software, this seems like a great trade to me.
(Disclosure: I work for a company with this business model, in part because I like working on open software)
When one develops using open-source software they have an obligation to follow the licenses. They also have a moral obligation to be respectful of the work upon which they’re building. And they have a social obligation to help improve that software where they can.
Those that develop on top of open-source have no obligation to give you their work for free.
You’d be surprised as to the amount of open-source contributions TIN drove towards Postgres, LLVM, and pgrx. And you’d be speechless at the amount of upstream work across all sorts of open-source PlanetScale does. Postgres 18.6, for example, is better for you today, in part, because of TIN. You’re welcome.
> You’re welcome.
I didn't say thank you and don't presume I would, Planetscale acting in their own self interest by improving postgres upstream isn't deserving of thanks, any more than Intel upstreaming a bunch of Linux kernel work is or myriad other examples.
Corporations acting in their self interest isn't worth giving thanks for, neither is the work of the people paid to do work on their behalf.
I don't expect my employer to thank me, I expect them to pay me, I don't expect users of software I was paid to write to thank me because I did it for the money not out of altruism towards those users.
It’s clear that we’re on opposite ends of open-source ideology.
Good luck out there! 2026 is wild times!!
You as well, and yes it's very "may you live in interesting times" at the moment everywhere.
Everything else aside
> Those that develop on top of open-source have no obligation to give you their work for free.
This is a pretty narrow view of open source.
Just as a simple example, GPLv3 and AGPLv3 are both considered Open Source and, depending on how you hold them, may obligate releasing work to customers essentially for free.
"...are both considered Open Source" << Free Software (GPL) is the O.G. - it's like saying "George Washington is considered to be an American President" or "The Beatles are considered to be a pop/rock band" or "water is considered to be..."
I remember so well when "Open Source" branding started with Bruce Perens - all the business arguments. When you need a database, someone ELSE'S business decisions (how THEY are going to make money) never end up helping YOU. If you use proprietary extensions and become dependent on them, you inevitably will get BURNED when their business needs diverge from your needs. - They close shop - They refuse to interop with something you need - They demand that you obey their arcane rules - They rug-pull - They get hacked as only they can - They lie to you - They stab you in the back
I thought I was clear:
> When one develops using open-source software they have an obligation to follow the licenses.
Obviously what follows from one of those licenses is what you say.
I’m glad we agree!
> When one develops using open-source software they have an obligation to follow the licenses.
Agreed!
> They also have a moral obligation to be respectful of the work upon which they’re building.
Respectfully, completely disagree!
Open source is a license, not a moral framework. Your only duty is to abide by the license. If you want to enforce that everything built on top of a particular open source package also be open source, that belongs in the license.
> Your only duty is to abide by the license.
We live in a society. I don’t see any harm in believing we ought to be respectful of the work. That’s all I’m saying. You don’t have to be, but it’s a better world if you are.
> If you want to enforce that everything built on top of a particular open source package also be open source, that belongs in the license.
Yes! I fully agree. The one that creates the thing is the one that gets to choose its license.
Having opened-sourced some bit of work myself, that’s a very difficult decision.
The problem is that they don’t support bare metal. I’d love to use PlanetScale in our bare metal servers.
we support bare metal inside AWS, GCP, and very soon Azure
what do you call bare metal in my office?
On-prem.
Postgres does have pg_fts (tsvector/tsquery/tsrank) which is a quite sophisticated full text search package integrated with functional indexing and query optimization. Why would I use something vibecoded that isn't part of core Postgres instead?
Please read the Postgres manual. It has incredible built-in search capability.
If you mean tsvector/tsquery - https://www.postgresql.org/docs/9.6/textsearch-intro.html - it's very good, but it's missing an important feature: ranking based on the overall document collection.
PostgreSQL built-in FTS provides a score for each row based just on the data for that row.
Relevance algorithms like BM25 take overall corpus statistics into account. If you search for a bunch of words and some of them are less common than others in the overall set of documents, documents that match THOSE words will score higher than matches for other words in your search.
That's what all of these additional extensions are providing.
If you read deep into this, they claim much better performance than the built in search; they also imply that the built-in search is missing features they provide but don’t make clear which ones (I think it is just support in the same index for queries covering other conditions on other columns, because every other feature they claim seems to line up with the built in search features, which have been around for about 20 years.)
Over Postgres' FTS, TIN provides at least:
There's a lot we didn't cover in the announcement blog. I'm sure we'll do more as time goes on.As an aside, something I personally think is cool, and I suppose you can do this with Postgres' built-in `@@` too, is that you can use TIN's full query language (linked above) against any text datum. This is a valid query:
in other words, you don't need an index at all to use TIN's full query language against any text field in any query.The built in search can't do any scoring mechanism that involves corpus-wide stats, so things like tfidf and bm25 are right out. If you don't need that then great, but in my experience the results are much worse.
It has it.
Incredible? No.
I think it’s fantastic.
You want to use this thing instead?
...yes, obviously?
Interestingly enough SQLites FTS supports Lucene queries out of the box with great performance characteristics. IIRC only writes become pretty slow after a while. I’ve always wondered what exactly would prevent PostgreSQL from strapping that implementation into its own database. My experience with ts_query hasn’t been particularly rosy. It can be better than LIKE but only marginally so and at the cost of insane index sizes… If this extension becomes open source and we can test it out in the real world I’m sure there’s a sweet spot
SQLite FTS relies on shadow B-trees under single-writer locks. Postgres index access methods must map postings directly to physical ctid tuples, surviving MVCC visibility checks and heap tuple churn.
There's another aspect which none of the FTS search solutions for Postgres do well in my opinion: multi-language support.
For example this one: it doesn't mention support for CJK languages (meaning tokenization for e.g. Chinese will resolve to one token per character, which will technically work and give results, but is inefficient). Also word stemming (databases -> database) is also missing as far as I can see, so the kind of queries where you'd expect related words to show up will be missing. Just doing case-folding and accent-folding is a bit of a functional but bruteforce solution.
Ideally I'd want something that supports:
- language aware tokenization, with ability to define the language per record. Including stemming, etc. And have useful predefined configuration for common languages (e.g. the Postgres built in one is missing many languages).
- CJK support, tokenizing at word boundaries.
- Optional accent- and case-folding.
Most solutions just seem to assume English content, I have not found anything that does all of this yet.
I struggle with FTS inside SQL (SQLite and MSSQL). There is often a fairly significant impedance mismatch between the relational concerns and how the documents need to be stored.
I've always preferred to use SQL as the system of record and then build/maintain an external Lucene index. Do we think these integral FTS capabilities are at the point where a hybrid architecture doesn't make sense anymore? How much customization exists in this provider?
I’m one of TIN’s developers and if you google my username you’ll see I’ve been in this space for a long time.
The answer to your first question is simply: yes
As far as your second question, what customization do you need that you believe TIN or PlanetScale doesn’t provide? These are things we can do, with alacrity.
in my experience it’s pretty common to find big inverted indexes for text directly in the database - not necessarily large docs but certainly free text records in volume. using bm25 and unicode’s breakiterator is a very good way to build it. like putting lucene in the database basically - makes a lot of sense when the database is already large. places that bend over backwards to move search out of the db are usually trying to avoid having a very large db (and often end up with one anyway, getting the worst of both worlds)
They also end up with all the infrastructure and processes necessary to keep the external search system in sync, resync/reindex, pkey shipping back to their source of truth in queries, application-side joins and enrichment between both sources. It’s brutal.
Having everything in one place eliminates entire classes of development and especially operational problems.
I want to try Planetscale... but we're addicted to (and totally dependent on) Neon's branching model. They really got us hooked on that!
Please note possible name collision with PostGIS Triangulated Irregular Network (TIN) data type.
I suppose it all depends on the scale of your project, but I've had pretty good luck using both MySQL's and SQLite's FTS capabilities. Surprised to hear that Open Source champion Postgres didn't have up-to-snuff FTS up to now...?
It has FTS built in and has had it for a VERY long time.
Why is this necessary? we have built in full-text search in postgres?
it does what it says on the TIN
It’s just me or there are others who keep seeing these updates and think mongodb had all of this years ago?
Seriously so happy to be running our production stack on mongo.
Poe's Law strikes again.
Is this an ad? Postgres has had search for more than a decade.
> Postgres has had search for more than a decade.
More than two decades (it moved to core from contrib in version 8.3 in 2008, but it was available in contrib since 7.4 in 2003.)
This was already posted and ignored at https://news.ycombinator.com/item?id=49751888 so i’ll ask the same question: again:
I don’t see any github link, is this 21st century embrace, extend, extinguish ?
>is this 21st century embrace, extend, extinguish
I assume here you're talking about Amazon's modus operandi?
What’s funny is I worked with a company with planet in the name Who could really use a full text search that was great in the Postgres
What are the advantages of Tin over using ts_vector with gin and gist indexes?
From the benchmarks deep in the document, TIN is much faster than built in text search (tested against GIN, which is itself much faster than GiST for text search.)