How It Works

How Distillary works

The core insight

Every source has an argument structure — a thesis supported by a hierarchy of claims. Distillary makes that structure explicit and navigable.

The full pipeline

graph TD
    Source["Source text"] --> Split["Split ~20KB chunks"]
    Split --> Extract["16 haiku agents → claims"]
    Extract --> Dedupe["Deduplicate"]
    Dedupe --> Entities["Extract entities"]
    Entities --> Group["Opus: group into clusters"]
    Group --> Pyramid["Opus: build hierarchy"]
    Pyramid --> Link["Find connections"]
    Link --> Assemble["Assemble vault"]
    Assemble --> Doctor["Doctor: fix + suggest"]
    Doctor --> Bridge["Bridge to other sources"]
    Bridge --> Brain["Brain vault"]

Purple = haiku agents (fast, parallel). Dark = opus agents (deep reasoning).

The pyramid

graph TD
    L3["Root thesis — 1 note"] --> L2a["Cluster: Validation"]
    L3 --> L2b["Cluster: Metrics"]
    L3 --> L2c["Cluster: Pivoting"]
    L2a --> L1a["MVP tests assumptions"]
    L2a --> L1b["Early adopters first"]
    L1a --> L0a["Zappos shoe photos"]
    L1a --> L0b["Dropbox video MVP"]
    L1a --> L0c["Concierge MVP"]

Layer	What	Count
3	Root thesis	1
2	Clusters (chapter-level)	~8
1	Structure (arguments)	~50
0	Atoms (evidence)	~300

The note format

---
tags:
  - type/claim/atom
  - priority/core
  - certainty/argued
  - stance/endorsed
  - domain/methodology
  - role/argument
  - source/ries-lean-startup
kind: claim
layer: 0
proposition: "validated learning → measures progress"
source_ref: "Chapter 7: Measure"
---

6 tag dimensions on every claim

Tag	Question
priority/	How central to the argument?
certainty/	How solid is the evidence?
stance/	Does the author endorse this?
domain/	What field?
role/	Fact, argument, prediction, definition?
source/	Which source?

Cross-source bridges

graph LR
    LS["Lean Startup:<br/>Vanity Metrics"] --> Bridge["Bridge:<br/>False Signals"]
    MT["Mom Test:<br/>Compliments"] --> Bridge
    Bridge --> BL["Both perspectives<br/>+ all backlinks"]

One bridge page = complete cross-source answer

The bridge has aliases from both sources. Obsidian resolves [[Vanity Metrics]] AND [[Compliments]] to the same page.

Agent navigation

graph TD
    Q["Question"] --> D{"Type?"}
    D -->|Summary| P["Walk pyramid"]
    D -->|Concept| E["Entity backlinks"]
    D -->|Compare| B["Bridge page"]
    D -->|Explore| C["Follow link chain"]

2-3 fetches for a complete answer

agent.json → entity page → backlinks. Multi-source answer with citations in ~2,500 tokens.