Flagship
Librarian
Semantic search across the studio's booksAsk a question, get the passage. An instrument that returns text you can use, not a summary you have to trust.
- Statusstable
- Since2026
- Accessinstall
Currently workingActive epics
- Proof Layer (Contract Hardening + Reader V1)P10E024 — Fechar o Proof Layer ponta-a-ponta: Search → Evidence → citation.url → Reader → texto visível na posição correta com highlight. Contract Hardening (locator íntegro, proof determinístico) + Reader V1 (materializar o trecho exato, highlight 800ms centralizado, zero chrome). Contrato do locator é a peça central.
- Bindery Acquisitionbindery-acquisition — Tornar o Bindery um canal de aquisição confiável e mensurável: saber quanto ele realmente cobre, por que os livros ficam em `wanted`, e rodar batches de download sem intervenção manual — fechando o loop até o índice (intake).
- Indexing — Master Epicindexing-master — One door into indexing. There are four epics touching this machine's index and no single place that says what is true *now*, what is broken, and in what order to fix it. This epic holds the live measurement, the failure taxonomy, and the sequence — and delegates detail to the four.
- Librarianlibrarian-readme — A BYOB (Bring Your Own Books) semantic search for your personal library.
- Publication Postflight — repo landing + doc parity fixespublication-postflight — TBD
- Proof Layer (Contract Hardening + Reader V1)librarian_20260714-reader — Fechar o Proof Layer ponta-a-ponta: Search → Evidence → citation.url → Reader → texto visível na posição correta com highlight. Contract Hardening (locator íntegro, proof determinístico) + Reader V1 (materializar o trecho exato, highlight 800ms centralizado, zero chrome). Contrato do locator é a peça central.
Coming soonPlanning epics
- Book Identity & Durable Permalinkslibrarian_book_identity_permalinks — Permalinks duráveis apoiados em IDENTIDADE BIBLIOGRÁFICA PADRONIZADA — não reinventar a roda. ISO/OL/DOI já resolvem identidade de livro. O trabalho é mapear cada TIPO de fonte para o identificador canônico existente e preservá-lo no sistema.
- DB Optimization — Memória & Latência do Motor FAISSP10E007 — Otimizar o motor de busca vetorial FAISS do Librarian: reduzir footprint de memória (4.4 GiB RSS, maior consumidor do Colima), baixar latência, e consolidar a arquitetura de retrieval/persistência do índice. Epic HEAD que agrega os sub-epics de otimização de DB/search como notes.
- Evidence & Provenance ModelP10E008 — Define como uma afirmação mantém ligação verificável com sua origem. Camada central do Librarian: Evidence, Source, Locator, CitationTarget.
- Index Optimization — Batching, Caching, Progressive PublicationP10E011 — O indexador atual leva ~15h e deixa o corpus indisponível. Três problemas: embedding sem batch (25x slowdown), sem cache de embeddings (crash perde tudo), FAISS só no final (corpus offline). Consertar cada um.
- Librarian PlatformP10E014 — Definir e estabilizar o Librarian como uma plataforma microkernel: um núcleo mínimo (Core Service :8088) que hospeda capabilities como extensões plug-in (Retrieval, Reader, Indexer, Log, Users), compartilhando processo, porta e corpus, mas com contratos independentes. Cada módulo deve ter um contrato testável antes de crescer (architectural freeze 2026-07-29: Research Triage / Retrieval Engine / Corpus / Evidence-Reader).
- Librarian Reliability LayerP10E016 — Librarian must never falsely report itself as healthy when its knowledge retrieval capability is unavailable. Index rebuild must never make corpus unavailable.
- Research Strategies — First Mile TriageP10E025 — Formalizar o Librarian como uma plataforma modular com 4 módulos independentes: (1) Research Triage & Planning, (2) Indexing/Knowledge Engine, (3) Knowledge Corpus, (4) Evidence Experience/Last Mile. O conceito central é o ResearchPlan — um objeto que carrega intenção, contexto, estratégia, confiança e critérios de sucesso — produzido pelo Triage e consumido pelo Engine.
- DB Optimization — Memória & Latência do Motor FAISSlibrarian_20260811 — Otimizar o motor de busca vetorial FAISS do Librarian: reduzir footprint de memória (4.4 GiB RSS, maior consumidor do Colima), baixar latência, e consolidar a arquitetura de retrieval/persistência do índice. Epic HEAD que agrega os sub-epics de otimização de DB/search como notes.
- Evidence & Provenance Modellibrarian_20260717 — Define como uma afirmação mantém ligação verificável com sua origem. Camada central do Librarian: Evidence, Source, Locator, CitationTarget.
- Experience History & Analyticslibrarian_20260630 — Track how books are experienced via RAG citations e leitura, não só lidos. Inclui: log/biografia de uso (timeline, co-ocorrência), citation history, ranking de livros citados-nunca-lidos, analytics por cluster e auditabilidade. Absorve E002 usage-analytics e o epic Log (2026-08-11).
In the same way that a well-designed object reveals how it wants to be held, this product reveals how it wants to be used before anyone reads a word. That is the whole ambition: cognition that shows itself in the surface.
What it is
It began as an internal instrument and stayed one. The shape came from use, not from specification — every part that survived is a part someone reached for more than once. What follows is a description of that shape.
The instrument does one thing: it maps a question onto the passages that answer it. It does not summarize. A summary asks you to trust a voice you cannot check.
How it works
Three movements, in order. First the intake, where material arrives unformed. Then the reduction, where it is made small enough to hold. Finally the return, where it goes back out into the same world that produced it.
flowchart LR
A[intake] --> B[reduction]
B --> C{holds?}
C -->|yes| D[return]
C -->|no| BWhat the movements cost
Each movement has a price, and the price is not hidden.
- Intake takes time proportional to the shelf.
- Reduction throws away everything that does not answer.
- Return is the only step the reader sees.
What it refuses
- It refuses to invent a passage.
- It refuses to average two authors into one voice.
- It refuses to return a page number it did not read.
A passage it returned
The instrument exists because this problem is old and already named.
The law of the instrument: give a small knowledge, and things look uniform, give a large one, and they look varied.
The same claim, sharper
We can prevent or reduce identifier collisions by adding information about the namespace, the domain from which the names or identifiers are selected.
Reading the output
A result is a passage, a source, and a score. The score is not truth; it is proximity. Two passages can score alike and say opposite things.
Inline detail appears like librarian.search(query) in a sentence, and a command
looks like librarian index --rebuild. Both are set apart from prose; neither
competes with it.
result = librarian.search("delay tolerant routing", k=5)
for passage in result.passages:
print(passage.book_title, passage.score)
$ librarian index --rebuild
indexed 412 books in 9m12s
A value you supply yourself is written --shelf <path> and is set apart from the
instruction, so the two never read as one thing.
Notes
Status
Working, with the usual caveats. Nothing here asks for attention it has not earned, and nothing claims to be finished.
| Surface | State | Since |
|---|---|---|
| Search | working | 2026-08 |
| Index rebuild | working | 2026-09 |
| Multi-shelf | not started | — |
Press ⌘ + K to open the query field.
Cover art is reserved for this product; the default cover ships first.
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris.
Lorem heading
Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
flowchart LR
A[Librarian — architecture] --> B[ingest]
B --> C{stable?}
C -->|yes| D[flagship]
C -->|no| E[experimental]
Get it
Lorem ipsum — install path, requirements, what lands on day one.