distillr 0.9.30

Source-to-intelligence platform: turn YouTube, websites, and arXiv papers into a structured, reusable corpus with per-source insights, cross-source synthesis, and Deep Research reports.

Distill

Installed as distillr on PyPI; the CLI is distill.

Distill helps you get insights on the topics you care about — and keep them current. Point it at a research goal; it finds the papers, videos, sites, and posts worth reading, analyzes each into structured insights with source receipts, and synthesizes across them into a durable plain-Markdown corpus on your disk. You browse it in Obsidian, your agents read it as files or query it over MCP, and it refreshes on a cadence instead of going stale.

pip install distillr
distill papers "temporal knowledge graph" --topic tkg --limit 20

That one command searches arXiv, downloads 20 PDFs, extracts full text, runs structured analysis on each, and writes a cross-paper synthesis. For a 20-paper run like the example below, expect single-digit minutes and under a dollar in model spend on the grok-4.3 default. Terminal output during the run looks like this (illustrative run; see the labelled sample-output note below):

Papers: temporal knowledge graph
Topic: tkg | Selected papers: 20

  [1/20] Time is Not a Label: Continuous Phase Rotation for Temporal Knowledge
         Graphs and Agentic Memory
  [2/20] Inductive Reasoning for Temporal Knowledge Graphs with Emerging Entities
  ...

  6m 47s  ~$0.58 (391,278 in / 38,117 out)

  time_is_not_a_label_260411544_Paper.md     90.4 KB
  time_is_not_a_label_260411544_Insights.md   8.1 KB
  ...
  tkg_Paper_Synthesis.md  11.8 KB
  tkg_Corpus_Synthesis.md 10.5 KB

Where distill sits

Three kinds of tools orbit this space, and distill is deliberately none of them:

• Deep Research oracles (ChatGPT, Gemini, Perplexity) are excellent at one-shot answers — and the work evaporates after each session. No corpus, no receipts you can re-check, nothing that compounds. Distill is the engine under that pattern: every run leaves transcripts, extracted paper text, per-source insights, and cross-source synthesis on disk, refreshable on a cadence.
• Grounded notebooks (NotebookLM) keep a persistent corpus, but in a silo: you find and feed the sources by hand, and the corpus exports to Google Docs/Sheets only. Distill finds the sources against your goal, and the corpus is plain files you own.
• LLM-wiki maintainers (the post-Karpathy wave of agent-curated Markdown vaults) assume you already have the content and tidy it. Distill is the acquisition half they leave out — goal-aware discovery across papers, videos, sites, and X, transcript-grade capture, and provenance on every claim — producing exactly the kind of vault those tools maintain.
• Academic literature tools (Elicit, Semantic Scholar, scite, Consensus) are stronger for pure paper search, citation graphs, and systematic review. Distill treats papers as one source type inside a broader corpus that also holds talks, vendor docs, and posts.

The short version: those are report and search layers; distill is the corpus layer underneath repeated research — capture, per-source insights, cross-source synthesis, refresh, receipts. And plain Markdown is the substrate, not the moat: anyone can write Markdown. The moat is the acquisition-and-maintenance loop that fills it and keeps it current.

That matters when you are doing thesis work, competitive analysis, technical due diligence, or building a startup knowledge base — you can verify the receipts, watch how a topic evolves, query the same folder through MCP from Claude Desktop / Cursor / other agents, and open it in Obsidian, Logseq, VS Code, or plain filesystem search. Nothing is locked in anything.

What you get

One local library/ directory of plain Markdown. No database, no cloud lock-in, no proprietary format. Files use globally descriptive names plus YAML frontmatter so knowledge-base tools, Dataview-style plugins, and AI coding assistants can understand them without guessing from generic insights.md tabs.

Four source types, same pipeline shape (capture -> analyze -> synthesize -> report):

• YouTube — channels, topic searches, videos, Shorts
• Websites — vendor sites, research hubs, curated URL sets (browser-first crawl with PDF/embedded-video ingestion)
• arXiv papers — phrase-matched search, full-PDF extraction, structured per-paper insights, cross-paper synthesis
• X (Twitter) posts — via distill ingest <tweet-url>; uses the public syndication embed endpoint (no anti-bot scraping). When a tweet has a native video attachment, the audio is transcribed via local-first Whisper (faster-whisper on GPU/CPU, OpenAI Whisper as cloud fallback) with a vocabulary hint derived from the source metadata to keep proper nouns intact.

Plus an MCP server so AI assistants and agent systems can query the library directly.

Quick start

Distill runs on Linux, macOS, and Windows (Python 3.12+); local models run on consumer GPUs via Ollama or LM Studio.

pip install distillr
playwright install chromium     # for YouTube search + website capture
distill doctor                  # verify API keys + system health

Set two keys in .env (copy from .env.example):

XAI_API_KEY=xai-...             # Grok models
GEMINI_API_KEY=AIza...          # Gemini Deep Research (reports + briefings)

Or run locally with Ollama (no API keys needed for ingestion):

ollama pull qwen3.5:27b         # download recommended model for 24GB GPU
echo "DISTILL_PROVIDER=ollama" >> .env
distill doctor                  # verify local setup

Then try any of:

# Goal-aware cross-source discovery (papers + videos + curated sites, reranked against a goal)
distill discover "help an AI become a great music composer" --topic music --preview
distill discover --goal-file private/my-goal.md --topic research --yes
distill discover --goal-file private/agent365-goal.md --topic agent365 --site-seeds private/agent365_sites.json --site-limit 10 --preview

# Get smart on a YouTube topic, fast
distill latest "Microsoft Fabric best practices" --limit 10 --report

# Discover and ingest arXiv papers — expands the query, LLM-reranks candidates,
# picks the top N (use --preview to see the shortlist without ingesting)
distill papers "agent memory systems" --topic memory --limit 20
distill papers "agent memory systems" --topic memory --limit 20 --preview

# Distill a vendor/research site
distill site-batch configs/example_seeds.json --topic example --seed-only

The full command reference lives in docs/usage.md.

Mental model

library/
  └── topics/<topic>/
       ├── channels/<creator>/videos/<video>/
       │     ├── <video-slug>_Transcript.txt
       │     └── <video-slug>_Insights.md
       ├── sites/<hostname>/pages/<page>/
       │     ├── <page-slug>_Content.md
       │     └── <page-slug>_Insights.md
       ├── papers/<paper>/
       │     ├── <paper-slug>_Paper.md
       │     └── <paper-slug>_Insights.md
       ├── <topic>_Topic_Synthesis.md      # cross-source
       └── <topic>_Corpus_Synthesis.md     # mixed-source view

You build a topic library over time. Ingest once, refresh on a cadence, generate a report or briefing when you need one. Older insights.md-style libraries are still readable, but new Markdown writes use the stable knowledge-base naming scheme.

See docs/outputs.md for what every artifact contains.

Sample output

The excerpts below are synthetic examples: the file shapes, frontmatter fields, and section structure are exactly what distill writes, but the papers, authors, and numbers are invented for illustration. For a provenance-first tool that distinction matters, so it is stated. A real, unedited example corpus (6 papers on claim verification, $0.19 of analysis) ships in examples/.

A cross-paper <topic>_Paper_Synthesis.md (excerpt):

## Strongest Research Signals

- Append-only temporal representations improve long-horizon extrapolation:
  RoMem (arXiv:2604.11544), EST (arXiv:2602.12389v3), and CID-TKG converge on
  persistent or dual-view entity state over destructive overwriting, with
  consistent MRR/Hits@K gains on ICEWS and GDELT.

- Semantic gating scales better than manual relation tagging: RoMem's Semantic
  Speed Gate and EST's energy-barrier gate both learn relational volatility
  from text embeddings rather than schema tags…

Per-paper <paper-slug>_Insights.md excerpt (click to expand)

---
title: "Time is Not a Label: Continuous Phase Rotation for Temporal Knowledge Graphs"
type: "insights"
topic: "tkg"
source: "arxiv"
source_id: "2604.11544v1"
url: "https://arxiv.org/abs/2604.11544v1"
authors: ["Alice Example", "Bob Example"]
tags: ["distill/tkg", "source/arxiv", "cs.AI"]
synthesis_scope: "single-paper"
analyzed_by: grok-4.3
source_mode: full_pdf
---

### Core Contribution
1. Continuous functional rotation θ_r(τ) = s · α_r · τ · ω instead of discrete
   timestamp lookup tables. Zero-shot interpolation of unseen dates.
2. Semantic Speed Gate: MLP that reads only text embedding ϕ(r) and outputs α_r.
   Learns relational volatility from data.
3. Geometric shadowing in complex space: obsolete facts rotated out of phase so
   the correct fact outranks contradictions via the scoring function alone.

### Methods and Evidence
- On ICEWS05-15, RoMem-ChronoR reaches 72.6 MRR (vs vanilla ChronoR 68.4).
- Zero-shot domain transfer to FinTMMBench: 0.728 MRR, 0.673 R@5.
- All baselines use identical answer LLM and judge for fairness.

### Limits and Open Questions
- Computational cost at millions-of-facts scale is motivation but no latency,
  memory, or throughput numbers are reported.
- Gate pretrained only on ICEWS05-15 political events; generalization to
  highly ambiguous relations is not quantified.

For multi-topic literature reviews, stakeholder briefings, or agent grounding, distill research-brief (Gemini Deep Research, web-augmented) and distill synthesize (grok-4.3 single-call, corpus-only) take a user-written context file that shapes the output. See docs/usage.md#research-briefings-and-deep-synthesis.

Dashboard

distill                         # terminal home screen
distill serve                   # local web dashboard at http://127.0.0.1:8899

The terminal home screen shows tracked topics, channel and topic watches, recent runs, failures, and rolling spend. The web dashboard adds clickable drill-downs to per-topic, per-channel, and per-video views with rendered markdown, plus cost history and watchlist status. Both auto-refresh and read directly from library files — no database.

MCP server, and agent-discoverable directories

Distillr is built for two parallel agent-integration paths:

Path 1 — MCP (structured queries). Claude Desktop / Claude Code config:

{ "mcpServers": { "distill": { "command": "distill-mcp" } } }

Distill exposes 21 tools (a deliberately small surface, shrinking toward workflow-shaped tools — the JIT read layer returns ranked path/preview/score tuples with read_insight drill-down, never full payloads by default), plus 12 resources and 4 prompts. See docs/mcp.md for the list.

Path 2 — file system (the corpus IS the interface). When a coding agent cds into library/topics/<your-topic>/, the directory is plain Markdown with stable filenames and YAML frontmatter, so grep, cat, ls, and find are first-class query primitives — no schema to learn, no MCP setup required. Every topic directory (and the library root) ships auto-generated CLAUDE.md and AGENTS.md orientation files with identical content — CLAUDE.md for Claude Code, AGENTS.md for Codex, Cursor, Gemini CLI, and the 30+ tools on the cross-vendor AGENTS.md standard — so any agent that enters the directory gets oriented. This matches what Anthropic's Agent SDK material recommends for agent design: file system + composable tools as the substrate, with structured APIs layered on top when they help, not as the only entry point.

There's also a canonical Agent Skill at skills/distill-corpus/SKILL.md — one vendor-neutral file teaching an agent how to read the corpus and drive the CLI (drop it into ~/.claude/skills/ or ~/.agents/skills/).

Cost

On the grok-4.3 default ($1.25/$2.50 per 1M tokens), bulk video analysis runs ~$0.03/video and a full paper $0.03; Gemini Deep Research dominates paid reports ($2–3/report); distill synthesize is ~$0.20–0.40 for a multi-topic corpus pass. grok-4.3 is the cloud floor — xAI retired the cheaper fast tiers (grok-4-1-fast etc.) on 2026-05-15, and those slugs now redirect to grok-4.3 and bill at grok-4.3 rates (migration guide). The only cheaper path is running analysis on a local model (Ollama/LM Studio) — distill eval --models grok-4.3,<local-model> measures the cost × quality tradeoff over frozen fixtures and recommends the cheapest model that clears your quality bar before you switch. Every run logs actual vs estimated cost to cost_log.jsonl, and the pre-run estimate self-calibrates against that history; distill costs shows it. The estimator's goal is accuracy, not safe padding — a padded estimate discourages runs you'd happily pay for, so calibration error is tracked and shrunk over time.

Providers are adapters behind a workload router: grok + gemini are the calibrated cloud defaults, Ollama/LM Studio the local route, and Anthropic and OpenAI adapters ship in-tree (wireable, opt-in). Broader backends (AWS Bedrock, Microsoft Foundry) and plan-quota compute — routing batch analysis through agent CLIs your existing subscriptions already license (Claude, Codex, Gemini, and others), eval-gated for quality — are committed on the roadmap.

Full cost model in docs/cost.md.

Reliability and trust boundaries

What's enforced (every release clears the same CI gate): ~1,950 tests at 81% branch coverage (floor ratchets up-only toward the 1.0 ≥95% gate), ruff + import-linter dependency-direction contracts + pyright + bandit + pip-audit, pinned dependencies via a committed uv.lock, SHA-pinned Actions, and PEP 740 build provenance on every PyPI release. Default tests mock all LLM and network boundaries — contributors never burn API spend; live integration tests are marked and opt-in.

Trust boundaries, stated plainly: everything ingested (transcripts, pages, PDFs, tweets) is treated as untrusted input — injection-resistance rules are threaded through first- and second-hop prompts, the dashboard sanitizes rendered HTML, and MCP file access is confined to the library root. Distill never bypasses login walls, captchas, or anti-bot defenses. Known-fragile edge: YouTube extraction depends on yt-dlp, which churns with YouTube's countermeasures — failures degrade with messages, not corrupted corpora. Analysis output is LLM-generated and can err; provenance fields on every artifact exist so you can check receipts, and a write-time claim-verification gate is the next major milestone (roadmap). Full posture: docs/SECURITY.md and the security section of the roadmap.

Docs

• docs/usage.md — full command reference
• docs/invariants.md — design charter: what distill is, is not, and the rules that don't bend
• docs/architecture.md — data flow, 4-phase report pipeline, model routing, security hardening
• docs/outputs.md — what every artifact contains
• docs/cost.md — cost model, examples, guardrails
• docs/mcp.md — MCP tools, resources, prompts
• docs/migration-grok-4.3.md — Grok 4.3 migration guide (model retirement May 15, 2026)
• docs/briefing-contexts/TEMPLATE.md — starting point for --context-file prompts
• private/README.md — where personal/client-specific files go (git-ignored)

Roadmap and changelog

• docs/CHANGELOG.md — what shipped
• ROADMAP.md — what's next

Feature work is interleaved with recurring bug-hunt + harden passes — see the release rhythm note in the roadmap.

Contributing

See docs/CONTRIBUTING.md for dev setup, quality gates, and scope. Security disclosures go through docs/SECURITY.md.

License

MIT — see LICENSE.