Biblicus

Make your documents usable by your assistant, then decide later how you will search and retrieve them.

Roadmap

This document describes what we plan to build next.

If you are looking for runnable examples, see docs/demos.md.

If you are looking for what already exists, start with:

Principles

Completed foundations

These are the capability slices that already exist and have end-to-end behavior specifications.

Retrieval evaluation and datasets

Retrieval quality upgrades

Context pack policy surfaces

Extraction evaluation harness

This evaluation harness compares extraction approaches in a way that is measurable, repeatable, and useful for practical engineering decisions.

Corpus analysis tools

Lightweight analysis utilities summarize corpus themes and guide curation:

Sequence analysis (Markov analysis)

Goal: provide a sequence-oriented analysis backend for corpora where order matters (conversations, timelines, logs).

Deliverables:

Acceptance checks:

Text utilities

Small, reusable building blocks for transforming text in ways that are hard to do reliably with one-shot generation.

Deliverables:

Acceptance checks:

Tactus integration

Tactus uses Biblicus as an ordinary Python dependency (biblicus>=1.11.3,<2). It shipped as library imports, not as a Model Context Protocol (MCP) toolset.

What Tactus uses:

Tactus is a consumer. Biblicus must keep these imports stable within a major version. A change that breaks one of them is a breaking change and needs a major version bump.

Next: curation surface, consumer contract, and learned decisions

Papyrus, the automated newsroom, uses Biblicus for curation. It calls the biblicus command by subprocess. The next work makes that use first-class and moves learned decisions out of Biblicus.

Curation surface on main

Goal: land the curation commands Papyrus already depends on, so that every command it calls exists on main and is specified.

Deliverables:

Acceptance checks:

Consumer contract for Papyrus

Goal: give Papyrus one stable, typed way to call Biblicus, with no guessing about paths or output shapes.

Deliverables:

Acceptance checks:

Learned decisions move to Cyclotron

Goal: stop learning human judgments inside Biblicus. Cyclotron (AnthusAI/Cyclotron) owns one recurring yes or no decision that learns from human review, such as whether a source is relevant or whether a proposal should be accepted.

Deliverables:

Acceptance checks:

Later: alternate backends and hosting modes

Goal: broaden the backend surface while keeping the core predictable.

Deliverables:

Acceptance checks:

Deferred: corpus and extraction work

These are valuable, but intentionally not the near-term focus while retrieval becomes practical end to end.

In-memory corpus for ephemeral workflows

Goal: allow programmatic, temporary corpora that live in memory for short-lived agents or tests.

Deliverables:

Acceptance checks: