How-To Guide — Haystack Archive

Getting Started

Logging In

Open the link you were given in your web browser. It will usually look like https://your-machine-name.tail*****.ts.net.
Enter the username and password you were given.
After login, use Enter the Archive to open the Research workspace at /app.
If you need to report a bug, open /report. Operator review/export lives at /bugs.

Tip: If the page does not load, the server may be offline temporarily. Try again in a few minutes or contact the person who shared the link.

What Matters for Beta Testers

The main workspace for testing is the Research workspace at /app. Operator tools such as the Indexer, Retriever API, and index-browsing screens now live separately in the operator workspace and are not part of the normal beta flow.

The Research tab now follows a simple 3-step workflow: configure the search, ask your question, then review the source list.

3-Step Workflow

Step 1 — Research Configuration

The left sidebar controls what part of the archive is searched and which AI model answers.

Archive Collections shows which parts of the archive are currently indexed.
Decades and Year are the primary time filters.
Author, Publication, and Topic hint sit under advanced filters.
Provider, Model, and Answer depth stay visible in AI Settings.
Advanced AI Settings holds Base URL, API Key, Temperature, and Max Tokens.

Phone note: On iPhone-sized screens, the workspace is intentionally simplified. Secondary controls are hidden so the flow stays focused on scope, AI choice, chat, and then sources.

Step 2 — Ask

The top-center area is where you choose Semantic Search or Keyword Search, review the current scope summary, and continue the conversation.

The running conversation stays visible above the question box.
The question box clears when you submit, but the question is added to chat history.
Search runs the query, Stop interrupts a long run, and New starts a fresh conversation.

Step 3 — Review

The answer stays in the chat. The lower review area is now focused on the current source gallery.

Sort By changes the order of the current results without rerunning the search.
Citations in the chat jump to the matching source in the gallery.
Helpful and Wrong sit directly under the chat so feedback is easy to give.
Saved Sources is currently deferred from the beta UI while the save/export workflow is being redesigned.

Search Modes

Mode	When to Use It
Semantic Search	The default meaning-based mode. Best for ideas, themes, arguments, and natural-language research questions.
Keyword Search	Best for exact names, phrases, titles, Bible references, publication names, or wording you expect to appear literally.

Tip: If you are unsure, stay in Semantic Search. Switch to Keyword Search when you want literal matches.

Examples of good semantic questions:

How did Adventists describe the Sabbath in the 1880s?
What arguments were used for health reform?
How were camp meetings discussed in the 1860s?

Examples of good keyword searches:

ten virgins
Uriah Smith
Review and Herald
2300 days

Narrow the Search

Filters work before retrieval. They shrink the corpus first, which often improves both relevance and speed.

Control	What It Does	Best Use
Decades	Limit the search to one or more decades	Best first filter for historical questions
Year	Limit the search to one specific year	Use when you know the exact year
Publication	Limit to one publication or periodical	Great for focused publication studies
Author	Limit to a specific author field	Useful for writer-specific questions
Topic hint	Add a soft nudge for the subject area	Optional; usually leave blank unless needed

Tip: For most historical work, start with Decades. It narrows the archive quickly without forcing the search to be too literal.

Archive scope note: archive choices currently come from the live collection roots, so the main archive filters you may see are Millerite and Pioneer Library. Older references to SDA now map to Pioneer Library.

Answer Depth

Answer depth changes how much evidence the system tries to gather and synthesize before answering.

Setting	What It Means
Quick	Faster, lighter evidence pass. Good for quick checks and simple lookups.
Balanced (recommended)	The best default for most beta testing and normal research work. Keeps the normal semantic path tighter and faster than Deep.
Deep	Slower, broader evidence pass. Reviews more sources and turns on the heavier semantic-analysis path automatically.

Important: Deep is the current way to request the heavier semantic-analysis path. There is no longer a separate main-screen Deep Analysis toggle. For broad concept questions, the normal Balanced flow can still use semantic expansion when the app detects that a wider concept-recall search is likely to help.

Reviewing Sources

Read the Answer in Chat

The answer now lives only in the chat window. This keeps the conversation history readable and makes follow-up questions feel natural.

Numbered citations like [1] and [2] refer to the current source list in Step 3.

Use the Current Sources List

The source gallery in Step 3 is the main evidence-review surface. A source card typically shows:

title or source heading
publication, author, date, and page details
snippet text
relevance ordering based on the current sort setting

Use Sort By to reshape the current source set. This does not rerun retrieval; it only reorders the results already returned.

Clicking the source thumbnail, the source card itself, or Open PDF should all open the same source in a browser tab. If a file downloads instead, report it through /report and include your browser/device.

Feedback

Helpful marks a response that worked well.
Wrong marks a response that missed the mark.

Important: Always treat the source cards as the primary evidence. The AI answer is helpful, but the archive sources remain the authoritative material.

Follow-Up Questions

You can continue asking questions in the same conversation. The app keeps recent context so follow-ups can build on the previous answer and source list.

Examples of useful follow-ups:

What about in the 1890s?
Did anyone disagree with that view?
Can you focus on Waggoner instead?

Click New when you want a fresh conversation rather than a continuation.

If the page refreshes unexpectedly, the latest chat and source slate should now restore automatically in the same browser.

Bug Reports

Use /report whenever the archive behaves strangely, refreshes unexpectedly, opens the wrong source, or loses your place.

The form remembers your email in the browser.
It automatically attaches your latest browser research session.
That attached snapshot includes recent chat turns, the latest question, current scope summary, status, and the current result set.

Advanced Tools

The main beta path does not require these sections, but they remain available under Advanced Tools.

Saved Searches for presets
Recent Sessions for reloading past runs
Technical Details for trace and runtime diagnostics
Current Search Export for exporting the current result slate
Force Deep Analysis as an operator override

In Technical Details, you may see retrieval-runtime information such as whether semantic search used sentence_transformers or fastembed, which device was used, and whether reranking ran on CUDA.

Tips

Start with Semantic Search. It is the best general mode for concept-driven research.
Use Decades early. It is one of the strongest ways to reduce noise in a historical corpus.
Switch to Keyword Search for exact wording. Names, titles, and set phrases often work better there.
Leave Answer depth on Balanced unless you intentionally want a faster check or a broader investigation.
Use follow-up questions instead of rewriting every query from scratch.
Review the source cards before trusting a strong-looking summary.

Troubleshooting

Problem	What to Try
Page does not load	The server may be offline. Try again shortly, or contact the operator.
Search returns no useful results	Broaden the query, remove filters, or switch between Semantic Search and Keyword Search. For broad concept questions, Balanced can still work well before you jump to Deep.
The AI says it cannot answer	Try a broader phrasing, reduce filters, or move from Keyword Search back to Semantic Search.
Results feel too broad	Add a decade, year, author, or publication filter before searching again.
Search is slow	Try Balanced instead of Deep, or narrow the corpus with Decades, Year, Author, or Publication.
Provider/model issues	Check AI Settings first. Provider and Model stay visible, while connection details live under Advanced AI Settings.
The page refreshed and my work disappeared	Refresh once more and let the session restore. If it still looks wrong, submit a report at `/report`.

Glossary

Term	Plain Meaning
Semantic Search	Meaning-based retrieval for ideas, themes, and natural-language research questions
Keyword Search	Literal matching for exact words, names, titles, or phrases
Source card	A result showing archive metadata plus a relevant passage snippet
Answer depth	How much evidence the app reviews before answering
Current scope	The plain-language summary of mode, filters, and AI settings shown above the composer
Technical Details	The trace area for diagnostics, retrieval/runtime notes, and other operator-facing details
Preset	A saved search setup you can reload later

Enter the Archive

← Back to Home