Theoria

Retrieval-augmented generation

The model can call dedicated tools to find and read material before composing an answer. Tool use varies with the question and the available evidence.

01
Question
Your question, conversation context, and chosen reading guide the answer.
02
Retrieve
Tools return passages, doctrinal texts, textual witnesses, or dated historical records.
03
Compose
The model synthesizes the retrieved material. Answer text streams into the conversation.
04
Inspect
Open a citation to read its source. Dated answers can link to Chronicle for context.

Retrieval engine

SQLite holds the indexed corpus. Search combines FTS5 keyword results with embedding similarity using reciprocal rank fusion. Direct reference lookups and specialized tools provide other routes into the texts.

Answer service

A Python / Starlette service connects the model to the corpus through the Model Context Protocol (MCP). The browser receives server-sent events (SSE); model credentials and MCP connections stay on the server.

Execution trace

Agent activity shows observed runtime phases, MCP calls, returned or failed outcomes, and elapsed time. It opens during generation and collapses after completion; expand it to inspect the recorded calls.

Historical context

When background helps explain the answer, the model is instructed to retrieve dated evidence and place a cited Historical context section at the end. Historical setting remains distinct from doctrinal authority.

Response animation

A small WebGL simulation beside the input follows request events: gathering on send, forming when answer text arrives, then settling and unwinding after the answer is saved. It makes no model calls and does not depict the model’s internal reasoning. It respects your device’s reduced-motion preference.

Retrieval supplies evidence; the model still interprets it. A citation lets you check a claim, but does not guarantee that the claim is correct.

A reading sets the interpretive perspective.

Each configured reading pairs an editorial profile with a corpus connection. It determines which standards lead the answer and which are presented as comparisons. Existing inquiries keep their original reading; switching starts a new draft.

Source boundaries

Coverage is selective and varies by reading. Ask for a named source and inspect the cited text.

Check the text

Citation chips open a source panel without leaving the answer. Compare the excerpt, reference, and attribution with the claim being made.

Explore the date

When citations resolve to dated records, “Explore this answer in Chronicle” opens a separate timeline with the answer’s context and a link back. Nearby records supply context, not additional proof.

Cologne Cathedral particle façade: adapted from © Raimond Spekking, CC BY-SA 4.0. A photo-derived west façade with shallow procedural relief, not a surveyed 3D model. Point-data provenance · License and adaptations.

What is saved, and what to check.

Browser session: No account is required. A signed, HttpOnly cookie identifies your browser session; the server checks that session before allowing access to an inquiry.
Stored inquiries: Opening Theoria, choosing a reading, or starting a new inquiry creates an unsaved draft. Switching between chat and Chronicle keeps your unsent question and view state in this tab’s session storage. Submitting the first question saves the inquiry on the server. History is available to anyone using that browser session; it is not an account or a device backup.
Model processing: Your question and the context used to answer it are processed by the configured model provider. Browser-scoped history does not mean the conversation stays on your device.
Evidence limits: The library is selective. Search can miss relevant passages, and generated answers can misread or misattribute them. Inspect citations, distinguish quotation from interpretation, and ask where traditions disagree.

Theoria is a tool for study. It does not speak on behalf of a church or replace pastoral counsel.

New inquiry

From question to source.