Loading today’s stories
Build1 distinct publisher3 min readPublished
The author of Doco argues that browse, search, cite and safe-edit operations over a maintained corpus do most of the work teams expect from embeddings, and that the pipeline adds a copy which drifts from the source.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Follow any of these and your For You feed starts watching them — no settings page required.
invest
Google's bank agent can read FactSet. The licence, not the model, decides how much.1 distinct publisher
build
When customers build their own features, your permission model becomes the product surface1 distinct publisher
build
A Notion agent that dies after each request, and the debugging error that broke version two1 distinct publisher
build
Three chunking revisions, RRF and a reranker, all shipped without one retrieval number1 distinct publisher
Start with the request about the on-call owner. It needs structured mutation and concurrency protection [7], and neither of those is a retrieval feature. The pipeline runs one direction, source to chunks to embeddings to index [16], so the edit still has to land on the source, with something holding the line against a concurrent write.
That is why the operation worth grading a knowledge base on is the last one in the post's list: patch a bounded unit against the version that was read [10]. That is compare-and-swap on a document block. A corpus that cannot express "this block, at this version" cannot support a safe agent edit, and no amount of embedding gets you there. None of the seven operations names an embedding model at all [23].
The drift list is the cost argument [17]. Every item on it is an identity problem: the chunk gets an address the user never sees, while the source already has one the user does. The post's advice is to preserve stable blocks, heading paths, versions and search where they already exist, before inventing another identity layer [18]. Two addresses for the same paragraph is one more than anyone volunteered to maintain.
Now the part that has to be tested before it transfers. Doco uses structured full-text search rather than claiming universal semantic search [13], built on SQLite's FTS5, which the official documentation describes as a virtual table with phrase, prefix, NEAR and boolean query support [14]. The author's position is that this is often enough for known policies, identifiers, names and operational terms [15]. What has to be true for that to hold on your corpus: your users have to search with roughly the vocabulary the documents use. Boolean and prefix matching finds the term the document uses, not the term the user guessed. When that misses, it misses quietly, which is why the post insists that no match in the first page must not be reported as no answer in the knowledge base [21]. The third request in the set, comparing every customer interview about onboarding confusion, is where ranking, clustering or semantic retrieval genuinely helps [8]. Ten million mixed-format documents is the case for a serious pipeline [9], and the original Lewis et al. framing is about generating from retrieved non-parametric memory for exactly those knowledge-intensive tasks [4].
Portability rests on MCP, which separates resources from tools so a server can expose contextual data and executable operations through a negotiated interface [20]. That is what lets the staged loop, list then outline then search then read then verify then answer, run under different agent clients without the knowledge base pretending to be a chat model [19].
The check I would run before funding an embedding pipeline is cheap. Ask the current system for the same block address twice, then submit a patch that should be rejected because the version moved. If either one fails, the hygiene work is the project, and the pipeline is being hired to cover for it [1].
Ranked by verification strength, evidence, and original report placement.
The original RAG paper by Lewis et al. combines a generator's parametric memory with retrieved non-parametric memory for knowledge-intensive tasks, and is especially relevant when the system must rank evidence from a large corpus and generate an answer from it.
The post considers four requests: what is the current refund window, show me the policy; update the on-call owner in the release checklist; compare every customer interview about onboarding confusion; answer arbitrary questions across ten million mixed-format documents.
The first request needs search, a canonical source, and a citation.
The second request needs structured mutation and concurrency protection.
The third request may benefit from ranking, clustering, or semantic retrieval.
The fourth request is a strong candidate for a serious retrieval pipeline.
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 30, 2026
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Checkable references, untested conclusion
The verifiable material here is thin but real: FTS5's query features, MCP's resources-versus-tools split and the Lewis et al. framing are all things a reader can go and confirm — though none of those documents sits in our coverage, only the author's summary of them. Everything the headline actually asserts, that lexical search over block-addressed documents is sufficient for ordinary workplace agents, is reasoning. The post says the important comparison is empirical and then does not run it.
No uptake to measure
Nothing in this story tracks use. The single usage fact is the author describing his own system's current retrieval method — no users, no deployments, no corpus sizes, no retrieval quality on any real workspace. A self-description of an in-progress project is not adoption, and we decline to score it as one.
A title's worth of overstatement
"Your AI agent does not need RAG" is a stronger sentence than anything in the body, which lists six conditions under which RAG is the right architecture and closes by telling readers to measure. Against that, the restraint is genuine: no semantic search is claimed for Doco, no numbers are dressed up as results, and the failure modes are hedged with "may". The gap is mostly headline, plus one unmeasured sufficiency claim doing more work than its support allows.
The architecture that matches the product
The seven operations demanded of a "readable knowledge base" are the operations Doco provides, and the one technique Doco does not have — embeddings — is the one the piece argues most teams do not need. That alignment is worth naming. So is the fact that the author names it himself, in the third paragraph, before making the FTS5 case. Disclosure changes what a reader should do with the argument; it does not change the argument's shape.
Clear on what is said, blind on whether it works
We can be confident about the contents of this argument, about the disclosure behind it, and about what Doco does today, because all three are stated plainly and consistently. We have no footing at all for the question a reader actually has — whether full-text search over stable block addresses beats a vector stack for the incident lookup, the policy citation, the one-paragraph edit — because a single author with a declared interest is the whole record.