# Winston — Historical Intelligence Archive (Full Context)

> Winston is an AI-powered research interface over the Focal Point Publications (FPP) archive: 153,000+ embedded document chunks spanning books, diaries, trial transcripts, expert reports, correspondence, signals intelligence, and legal documents related to the work of historian David Irving.

## About

- **URL**: https://ask.winston.study
- **Archive source**: [Focal Point Publications](https://fpp.co.uk)
- **Document count**: 153,216 embedded chunks across 500+ source documents
- **Built by**: [irving.studio](https://irving.studio)
- **Technology**: RAG (Retrieval-Augmented Generation) with ChromaDB vector database, Ollama for local inference, Anthropic Claude API for cloud models
- **Embedding model**: nomic-embed-text (768 dimensions)
- **Cross-encoder reranking**: ms-marco-MiniLM-L-6-v2

## Architecture

Winston operates as a hybrid search + generation system:

### Search Pipeline
1. **Vector search**: User query is embedded via nomic-embed-text and matched against 153,216 document chunks in ChromaDB using cosine similarity
2. **Cross-encoder reranking**: Top candidates are reranked using ms-marco-MiniLM for relevance scoring
3. **Source diversity enforcement**: Maximum 2 chunks per source document to prevent single-document domination
4. **Person query filtering**: For "who is X" questions, sources are filtered to those mentioning the person's name
5. **Supplementary search**: If insufficient source types are found, additional targeted searches fill gaps
6. **Meilisearch fallback**: Full-text search across two indexes (fpp_full, fpp_posts) for keyword matches

### Generation Pipeline
1. Sources are truncated to 2,500 characters each to fit within context window limits
2. Top 4 sources are tagged as "MOST RELEVANT" in the user prompt
3. System prompt enforces strict anti-hallucination rules: answer ONLY from provided sources
4. Model generates a response with inline source citations [Source 1], [Source 2], etc.
5. Response is streamed via Server-Sent Events (SSE)

### Available Models
- **Gemma 3 12B** (Local) — Free, runs on local NVIDIA GPU, 8192 token context window
- **Claude Sonnet 4** (API) — ~$0.02/query, Anthropic, superior answer quality
- **Claude Opus 4** (API) — ~$0.07/query, Anthropic, highest quality
- **Claude Haiku 3.5** (API) — ~$0.003/query, Anthropic, fastest

## Document Types in the Archive

### Books and Manuscripts
Major works including:
- *Churchill's War* (Volumes 1–3)
- *The Destruction of Dresden*
- *Hitler's War* (1977 and 1991 editions)
- *The Rise and Fall of the Luftwaffe: The Life of Field Marshal Erhard Milch*
- *Rommel: The Trail of the Fox*
- *Göring: A Biography*
- *The War Between the Generals*
- *Uprising! Hungary 1956*
- *Hess: The Missing Years*
- *Nuremberg: The Last Battle*
- *Banged Up* (memoir)
- And others

### Trial Transcripts
- **Irving v Lipstadt** (2000) — Full trial transcripts, all 32 days
- **Nuremberg Trials** — Selected transcripts and documents

### Personal Diaries
- David Irving's personal diaries spanning 1978–2007
- Contains research notes, travel records, meeting notes, personal reflections

### Expert Reports
- **Evans Report** — Prof. Richard Evans' expert report for the Irving v Lipstadt trial
- **Longerich Report** — Prof. Peter Longerich's expert report
- Other expert submissions and witness statements

### Signals Intelligence
- **CSDIC transcripts** — Combined Services Detailed Interrogation Centre recordings of captured German generals
- **Ultra decrypts** — References to Bletchley Park decrypted communications
- Intelligence assessments and analyses

### Correspondence
- Letters to and from David Irving
- Reader mail from fpp.co.uk
- Publisher and editor correspondence

### Articles and Posts
- Website articles from fpp.co.uk (8,878+ published posts)
- Press releases, news commentary, research notes

## API Reference

### List Models
```
GET /api/models
```
Returns available models with id, name, type, and cost.

### Streaming Query
```
POST /api/query/stream
Content-Type: application/json

{
  "question": "What happened at the bombing of Dresden?",
  "model": "sonnet",
  "session_id": "optional-session-id",
  "source_filter": "optional-type-filter"
}
```

Returns Server-Sent Events:
- `{"type": "sources", "sources": [...]}` — Retrieved source documents
- `{"type": "token", "text": "..."}` — Streaming response tokens
- `{"type": "done", "responseTime": 5.2}` — Completion event

### Non-Streaming Query
```
POST /api/query
Content-Type: application/json

{
  "question": "...",
  "model": "local"
}
```

Returns complete response with answer and sources.

### Health Check
```
GET /api/health
```

## Query Caching

Responses are cached by question + model combination. Repeat queries return instantly (0.0s).

## Important Notes

- **AI Research Tool**: All answers should be verified against primary sources. The system can make errors.
- **Source Grounding**: Answers are generated ONLY from retrieved archive sources. The system is instructed never to use external knowledge.
- **Archive Scope**: The archive reflects the published works, collected documents, and personal papers of David Irving. It is not a comprehensive historical database.
- **Anti-Hallucination**: Strict system prompt rules prevent the model from fabricating information. If sources don't contain relevant information, the system says so.
- **Source Citations**: Inline citations like [Source 1] reference the numbered sources returned with each response.

## Links

- [FPP Archive — fpp.co.uk](https://fpp.co.uk)
- [irving.studio](https://irving.studio)
- [Winston Terminal — ask.winston.study](https://ask.winston.study)

---
© Focal Point Publications. All rights reserved. Built by irving.studio.