Historical Persona AI — Glossary and Sources

Companion note to the conversation on building a text or speech historical-character interface grounded in a real person's writings and contemporary knowledge.

How to use this in Obsidian

  • Terms are written so you can [[wikilink]] them later (e.g. [[RAG]]).
  • Sources are grouped by topic. URLs are plain markdown links.
  • Callouts mark definitions vs. caveats.

Glossary

Core methods

RAG

Retrieval-Augmented Generation. Instead of asking the model to answer from memory alone, the system first searches a document collection you control, then feeds the retrieved passages into the prompt so the answer is grounded in those texts.

For a historical character this is the main accuracy tool: letters, speeches, and dated notes are retrieved before the model speaks. It does not by itself stop anachronism; you still need date filters and refusal rules.

Related: vector database, embedding, hybrid search, BM25, lorebook.

LoRA

Low-Rank Adaptation. A way to fine-tune a large model by training a small number of extra weights instead of rewriting the whole model. The result is a compact “adapter” you can load on top of the base model.

Use LoRA for voice and rhetorical habit (sentence rhythm, favorite metaphors). Do not expect LoRA alone to erase the model’s modern world knowledge.

Related: QLoRA, PEFT, fine-tuning, SFT.

QLoRA

Quantized LoRA. LoRA plus quantization (storing the frozen base model in 4-bit, typically). Same idea as LoRA, much lower VRAM. The usual method if you fine-tune an 8B–32B model on a single consumer GPU.

OCR

Optical Character Recognition. Software that turns photographed or scanned pages (PDF page images, newspaper scans, manuscript photos) into searchable text.

Primary-source pipelines almost always start here. Dirty OCR (long-s as f, broken line wraps, running headers) will pollute retrieval, so clean it before chunking.

LLM

Large Language Model. The generative engine (Claude, GPT-class, Gemini, Llama, Qwen, GLM, Kimi, etc.) that writes the character’s reply.

PEFT

Parameter-Efficient Fine-Tuning. Family of methods that adapt a model without full retraining. LoRA and QLoRA are the common members.

SFT

Supervised Fine-Tuning. Training on example pairs: user question → in-character answer. Better for a persona than dumping raw collected works into next-token pretraining, especially if you include “I would not know that” examples.

DPO

Direct Preference Optimization. A fine-tuning method that teaches the model to prefer one reply over another (e.g. historian-rated authentic vs. anachronistic). Used in some character-roleplay papers together with SFT.

RLHF

Reinforcement Learning from Human Feedback. Older / heavier cousin of DPO: a reward model scores answers, then the policy is optimized. Mentioned in some historical-chatbot writeups as a way to reward “sounds like Churchill in 1943” over “explains the internet.”

Memory, search, and prompting

embedding

A numeric vector that represents the meaning of a text chunk. Similar passages land near each other, which is how semantic search finds “that letter about the harvest” even if the user did not use those words.

vector database

Store for embeddings plus metadata (date, place, source type). Common tools: Chroma, Qdrant, LanceDB.

Combine semantic search (embeddings) with lexical search (BM25). Useful when the user quotes a proper name, a statute title, or a distinctive 19th-century phrase.

BM25

Best Matching 25. A classic keyword ranking function. Complements embeddings; does not replace them.

chunk / chunking

Splitting long documents into retrieval-sized pieces (a letter, a speech paragraph, a diary entry). For historical work, document-and-date-aware chunks beat naive “every 500 tokens.”

metadata

Fields attached to each chunk: source_date, place, source_type (letter / speech / diary / later biography), certainty, first_person. Date metadata is what lets you refuse posthumous knowledge.

lorebook

Also called World Info in SillyTavern. Keyword-triggered snippets inserted into the prompt only when those keys appear. Lighter than full RAG; good for stable facts (birth year, offices held, family names).

character card

A portable persona file (JSON or PNG-with-metadata, Character Card V2/V3). Holds description, personality, scenario, first message, example dialogue, and sometimes an embedded lorebook.

few-shot

Including a handful of example Q&A turns in the prompt so the model copies register and refusal style.

system prompt

The hidden instruction that defines identity, “today’s” date, ignorance rules, and citation habits.

guardrail

A second pass (rules or a small model) that flags anachronisms, modern brand names, post-death dates, or later scientific explanations, then forces a regenerate or a refusal.

hallucination

Fluent invention presented as fact. RAG reduces it; it does not eliminate it. Always keep a “show the retrieved passage” toggle while you are building.

knowledge cutoff / temporal bound

The last date the character is allowed to know — usually death date or a chosen “now” (e.g. 1863 Lincoln). Separate from the model’s training cutoff.

Training and hardware

fine-tuning

Any extra training after the base model exists. Full fine-tune = change all weights (expensive). LoRA / QLoRA = change a thin adapter.

pretraining

The original internet-scale training run. You will not redo this for one historical figure.

LoRA rank (r)

Size of the adapter. Common starting point: r = 16. Higher rank = more capacity and more risk of overfitting a small letter corpus.

DoRA

Weight-Decomposed Low-Rank Adaptation. A LoRA variant some 2026 toolchains default to.

Unsloth

Single-GPU fine-tuning toolkit; fast LoRA/QLoRA on consumer cards.

Axolotl

YAML-configured fine-tuning pipeline; better when you want reproducible multi-method runs (SFT, DPO, etc.).

LLaMA-Factory

Fine-tuning UI / toolkit covering many alignment methods.

VRAM

Video RAM. GPU memory. Determines whether a local model or a QLoRA job fits.

GPU / CPU / CUDA / MPS

Graphics Processing Unit / Central Processing Unit / NVIDIA’s compute platform / Apple Silicon GPU backend. Local inference and training care which you have.

quantization (4-bit, Q4, etc.)

Storing weights at lower precision so a bigger model fits in less VRAM, with some quality cost.

Speech interface

STT

Speech-to-Text (also ASR, Automatic Speech Recognition). Microphone → transcript. Examples: Whisper, Parakeet, Deepgram, AssemblyAI.

TTS

Text-to-Speech. Model reply → audio. Examples: ElevenLabs, Cartesia, Kokoro, Chatterbox, GPT-SoVITS, Fish Audio.

TTFA

Time To First Audio. Latency until speech starts. Matters for spoken conversation, not for a typed prototype.

WER

Word Error Rate. STT accuracy metric. Lower is better.

MOS

Mean Opinion Score. Human listening-quality rating, usually 1–5. Used in TTS papers.

voice cloning

Synthesizing speech that matches a reference recording. Legally and ethically the sharp edge for real historical figures; a period-plausible voice is usually safer than cloning a specific performance.

realtime / speech-to-speech

Pipelines that skip a separate text step (e.g. OpenAI Realtime). Convenient, harder to inspect for anachronism because you never see the intermediate prose.

NIL

Name, Image, and Likeness. The commercial identity that publicity-rights law protects.

right of publicity

State-level (in the U.S.) right to control commercial use of identity. Often post-mortem for a fixed term (examples discussed in sources: California commonly cited at 70 years, Texas 50, New York shorter / performer-focused). Not uniform nationally.

digital replica

Statutory term in newer AI bills and California A.B. 1836: a computer-generated, highly realistic voice or visual likeness. Commercial audiovisual use of a deceased personality’s replica may need estate consent; commentary, criticism, parody, scholarship, and some documentary/biographical uses are often carved out — check counsel, do not treat this note as legal advice.

public domain

Works whose copyright has expired (in the U.S., generally author death + 70 years for many works, with older rules for pre-1978 publications). Their letters may be free to index; a 2019 annotated biography usually is not.

Other acronyms that showed up in the research

Acronym Expansion Why it appeared
API Application Programming Interface How hosted models and TTS/STT are called
UI User Interface Chat frontends
PDF Portable Document Format Source scans
KG Knowledge Graph Some biography systems (e.g. AIstorian) index facts as graphs, not only passages
MEI Major Entity Identification Entity/coreference step in Fic2Bot
GDPR General Data Protection Regulation EU living-person data law; generally does not apply to the long-dead, but user chat logs still might
HIPAA Health Insurance Portability and Accountability Act Irrelevant unless you store health data; appeared only in voice-platform comparison tables
PNG Portable Network Graphics SillyTavern character cards often embed JSON in a PNG avatar

Suggested Obsidian map

[[Historical Persona AI — Glossary and Sources]]
    ├── [[RAG]]
    ├── [[LoRA]] / [[QLoRA]]
    ├── [[OCR]]
    ├── [[character card]]
    ├── [[lorebook]]
    └── Sources
         ├── Methods (RAG, episodic memory, role-play)
         ├── Tooling (frontends, fine-tune, voice)
         ├── Primary-text repositories
         └── Law and ethics

Sources

Collected from the research pass on 2026-09-19. Annotations say why the item matters for this project, not that every claim in the paper was independently re-verified.

Methods — RAG, persona, historical dialogue

Tooling — frontends, roleplay models, fine-tuning

Voice (STT / TTS)

Use these only after the text persona is stable.

Primary-text repositories (examples, not exhaustive)

Also search, once you pick a person: Founders Online, presidential libraries, collected-papers projects, Chronicling America, HathiTrust, national archives of the relevant country.

Label the reconstruction

Several of the persona papers argue that users should be told this is a text-grounded simulation, not the person. That is both an ethics practice and a practical defense when the model is wrong.

One-page stack reminder

  1. Closed corpus of dated primary texts (after OCR cleanup).
  2. RAG with date metadata — no live web in character mode.
  3. System prompt + guardrail for ignorance and anachronism.
  4. Optional LoRA / QLoRA on 500–2,000 grounded Q&A pairs for voice.
  5. STT / TTS only after the prose persona holds up under red-team questions.

See also