Seer: the personalized-retrieval
Part I — The Server · Chapter 1 of 9
This started on October 23, 2025, with a server called Seer.
Not an app, not an agent, not a wrapper around a chat completion — a server, because the thing I wanted to build could not live in a prompt. I wanted an assistant whose retrieval knew me: which of my documents actually matter to me, which ideas I return to, which sources I trust — learned from behavior, not declared in a settings pane. Everything else in this book descends from that decision, so this chapter is about what Seer is, and about the first honest look I took at it.
A server that reasons
Seer became the reasoning lane of the eventual system: the voice, the retrieval, and the identity authority. It owns every spoken word and all retrieval-augmented grounding; a client sends conversation messages and a retrieval scope, and receives prose back. It mediates authentication, and it returns something most systems don't: per-owner attribution spans — which characters of the reply were shaped by whose context, with influence weights. From the very beginning I wanted provenance to be a first-class output, not a log entry. That decision looks small here; it becomes load-bearing in chapter 9.
The interior was three named subsystems, and their names will recur: Sinatra, the personalization engine; Oracle, the peer-to-peer retrieval mesh; Gita, the attribution layer.
Sinatra: the feedback loop
Sinatra is the mechanism that makes retrieval personal. Every retrieval parks — lands in a turn. The user's next turns carry a reaction, and that reaction gets scored as sentiment. The sentiment trains a small gradient-boosted tree. The tree re-ranks the next retrieval. Park, sentiment, train, infer — a closed loop running across consecutive turns of ordinary conversation, with no explicit feedback UI at all.
The features are the unusual part. I borrowed MACD and stochastic indicators — trend and momentum instruments from market analysis — and computed them over engagement instead of price. In March 2026 I sat down and reviewed the whole architecture as coldly as I could manage, and the strengths half of that review is worth quoting as it was written:
Field note — 2026-03-06. "The Sinatra feedback loop is the most interesting part. Park → sentiment → train → infer across consecutive turns is a clean design. Using MACD and stochastic indicators as features for a GBT is unusual but defensible — those signals capture trend and momentum in engagement over time, which is exactly the right shape of the problem. […] Oracle's decoupling is well-done.
OracleDelegateandOracleTransportas separate protocols means the P2P logic, the database, and the wire format are genuinely independent. […] The trust scoring system on edges is self-correcting. Unreliable peers sink organically without any manual intervention."
Engagement with your own knowledge is a time series. It trends, it has momentum, it mean-reverts. Treating it that way meant the model's features captured how your interest is moving, not just where it pointed last.
Oracle, Gita, and the debts
Oracle was the ambitious one: a peer-to-peer mesh where specialized nodes answer each other's retrieval queries, with trust scores on edges that decay when a peer returns junk — so unreliable peers sink organically, without moderation. Its decoupling was genuinely good; its value was deferred, and the review said so: with one or two nodes, a mesh adds latency and complexity for marginal gain. The payoff needs many nodes with non-overlapping knowledge.
Gita computed attribution weights over those spans — a layer built ahead of its moment, and the review was blunt about exactly that.
And the review kept going, into the parts that were mine to fix. At review time: Sinatra needed a run of interactions before it would even generate features and several samples before it would train, so every new user's first conversations ran on default retrieval — a cold start documented nowhere. The relevance threshold was hardcoded in three places, silently coupling the whole fan-out behavior to one embedding model's distance distribution. The GBT retrained fully on every chat turn — sub-millisecond at that scale, but the interaction history had no cap, so the cost curve bent upward with every power user. And a managed backend served as both auth and logging, a single point of failure wearing two hats.
I'm including the debts because of what the review was, not just what it found. The takeaways I wrote down that day — improve search; rethink the retrain and what interaction history even means; make the relevance threshold dynamic to the dataset; defer the mesh but keep its patterns — became the actual roadmap. Chapters 2 and 3 are two of those takeaways, pursued to their conclusions.
The claim, stated
The review ended by naming the claim underneath everything, and what it would take to measure it:
Field note — 2026-03-06. "[…] personalized retrieval — not just better generation — is what makes an AI assistant feel like yours. […] Whether it actually works well in practice depends entirely on whether the GBT's features capture something real about the user's retrieval preferences, which is an empirical question you won't know until it's in real users' hands."
Read that against the industry backdrop of early 2026: model quality was commoditizing fast, and nearly everyone's "personalization" was a text file of remembered facts injected into a prompt. The claim was that the differentiation would live below generation — in infrastructure that learns, per person, what is worth retrieving at all.
One more thing survives from this chapter, and it's a method rather than a mechanism. That March review — adversarial, written down, dated — was the system observing itself honestly and feeding the observation back into design. Every later part of this stack has that shape: ledgers that record what retrieval actually did, receipts that record what execution actually changed, provenance axes that record how a fact is known. The habit started here.
The candid review was the first ambient instrument — the system was being built by watching itself honestly.
What survives into the paradigm: retrieval that learns the person.