Mira — Research Platform

Privacy Doctrine

Loading…

Atelier Settings

View

Display density

Adjust text size and spacing in the Atelier.

2

Show idle research prompts

When the Atelier composer is empty, cycle through a research question drawn from your indexed library.

UI · Motion & Reveal

Coordinated reveal animations

Fade research panels in with a gentle stagger when a query starts. Off — everything appears instantly.

Animation tempo

Speed of the reveal sequence. Snap is closest to instant; calm is deliberate.

Idle prompt cadence

Seconds between idle research prompts in the empty composer. Recommend 25–45.

30 s

Setup

Re-run Setup

Reopen the setup wizard — how your research is powered, your university library, and your Mira Library folder. Change any of it any time.

Developer Mode

Embedded Claw Code terminal in the Atelier (Alt+Shift+D).

OFF

Credits

Research

Research depth

How hard a single researcher digs into your documents: higher means more searches and a more thorough answer, but it takes longer and costs more.

12

Research context (per collection)

Describe your research question, key figures, and what counts as an interesting finding. Stored per collection — each collection has its own context. Also editable one click from the Atelier composer.

Extended Thinking

Enable extended thinking

Some models can pause to reason through a hard question before they answer. This can sharpen difficult analysis, but it is slower and uses more credits.

Reading Room

Search Views — choose the question you’re asking

The Reading Room offers four views over your library. They use the same query and the same scope, but present the answer differently. Pick the view that matches the question you’re actually asking.

Concordance: Select the Concordance view to see every single sentence in which your word or search term appears. The matching sentences are gathered together and shown to you as a series of “citation cards.” A citation card is a small panel that represents one result: it names the source the sentence came from (its title, author, year, and which passage of the document it is), it shows you the full sentence with your search term highlighted, and it gives you an “Open passage” button that opens that source in the Reader so you can read the surrounding text. Use this view when you want to read every use of a word in the context of its own sentence. By default Mira matches whole words only, so a search for “gender” will not also match “engendered”; if you would like those partial matches as well, you can turn off whole-word matching in Advanced Search.
Dispersion: Select the Dispersion view to see where in each document your term appears, rather than only how often. Mira draws one horizontal strip for each document and places a vertical tick mark at every point in the text where your term occurs, so you can tell at a glance whether the hits are spread throughout the work or clustered in one place. The strips are ordered by the number of hits by default, and you can switch them to be ordered by how evenly the term is spread. Beside each strip is a “DP score.” DP is short for Deviation of Proportions, a measure devised by the linguist Stefan Gries: a score of 0 means the term is spread perfectly evenly across the document, and a score of 1 means it is concentrated in a single spot. Use this view to tell the difference between a theme that runs through an entire work and a word that is mentioned only once.
Frequency: Select the Frequency view to count how often your term appears, grouped by document, by collection, or by year. Alongside the raw count, Mira shows a “per-10,000-characters” figure, which is the count adjusted for how long each document is, so that a short article and a long book can be compared fairly. You can click any row to jump straight into the Concordance view, narrowed to just that document, collection, or year.
Relevance: Select the Relevance view when you are looking for an idea rather than one exact word and you want the best matches shown first. Mira ranks the passages by how well they match what you asked for, using one of three methods that you can choose in Advanced Search: keyword matching (a long-established method called BM25 that gives more weight to words that are rare and that appear often in a passage), meaning-based matching (which compares the idea behind your query to the idea in each passage, even when the exact words differ), or a blend of the two. For example, a Relevance search for “colonial intimacy” will surface passages about that idea even if they never use those exact words — which is different from the Concordance view, where you would only find passages that literally contain the word “colonial.” When the Relevance view is showing, a “Rank by” choice appears just under the view tabs, where you can switch between Keyword (your exact words), Smart Search (a blend of both, and the best choice for most searches), and Semantic (meaning only). This choice only affects the Relevance view.

The Concordance, Dispersion, and Frequency views are exhaustive: they return every single match, leaving nothing out. The Relevance view works differently — it is ranked, which means it sorts the matches from strongest to weakest and shows you the best ones first. (Researchers sometimes call this returning the “top-K” results, where K is simply however many of the top results you choose to see.) Because it ranks, the Relevance view may also leave out very weak matches so the strongest ones stand out, whereas the exhaustive views never leave anything out.

How your collections are searched

When you add documents to Mira, the application breaks each document into short passages and converts them into a matrix — a large table of numbers where each row represents one passage and each column captures one dimension of that passage’s meaning. (This preparation happens in Resource Management.) When you use Mira’s meaning-based search here in the Reading Room — the Semantic or Smart Search methods — or when you run research in the Atelier, your search queries and prompts are converted into rows of numbers in the same way. Mira’s search tools then compare your query’s row against every passage row in the matrix to find the closest matches. The passages that match most closely appear first in your results; passages with weaker matches appear further down, or not at all.

Your collection, represented as a matrix: dimension → d1 d2 d3 d4 … d_n passage 1 │ 0.12 0.84 −0.33 0.07 … 0.41 │ passage 2 │ 0.55 −0.02 0.71 0.39 … −0.18 │ passage 3 │ 0.08 0.79 −0.28 0.11 … 0.38 │ ⋮ │ ⋮ ⋮ ⋮ ⋮ ▿ ⋮ │ passage n │ −0.61 0.23 0.44 −0.09 … 0.52 │ your query │ 0.10 0.81 −0.30 0.05 … 0.40 │ ← compared against every row above Result: passages 1 & 3 are nearest to your query → they appear first in your results.

Each number in the matrix represents how strongly a passage relates to one abstract dimension of meaning. No single number is human-readable on its own, but taken together as a row, the hundreds of numbers form a unique fingerprint for that passage’s content. When two rows have similar numbers in the same columns, the passages are about similar topics — even if they use completely different words.

Learn more: What is Vector Embedding? · What is a Vector Database? · What are LLMs?

Retrieval Methods

Keyword (BM25)

What it does: BM25 searches for the exact words you type. It looks through every passage in your collections and asks two questions: does this passage contain my search terms? and how important are those terms? A word like “dialectical” that only appears in a handful of passages is treated as a much stronger signal than a common word like “the.” If your search term appears many times in a short passage, that passage is judged more relevant and appears earlier in your results — closer to the top of the list. A long passage where the same term appears just once, buried among thousands of other words, is judged less relevant and appears further down.

When to use it: Use BM25 when you are looking for specific words, names, or phrases that you know appear in the text. For example, searching for “Feuerbach” or “rate of profit” — where the exact term matters. It is the most predictable method: if the word is in your collections, BM25 will find it. It does not require the matrix of numerical representations described above; it works directly with the words in your documents.

Why it works: BM25 rewards rarity (uncommon words count more), repetition (a word appearing several times signals relevance), and brevity (a focused passage beats a rambling one). These three principles, combined, make it very good at finding exactly what you asked for — but it will miss passages that discuss the same idea using different vocabulary.

Mathematical notation

Best Match 25 — a probabilistic ranking function that scores documents by term frequency, weighted by inverse document frequency and normalized by document length. score(D, Q) = ∑_i=1ⁿ IDF(q_i) · f(q_i, D) · (k₁ + 1) ⁄ [ f(q_i, D) + k₁ · (1 − b + b · |D| / avgdl) ]

Q = the query, composed of terms q₁ … q_n
D = a document being scored; |D| = its length in tokens
f(q_i, D) = raw frequency of term q_i in document D
IDF(q_i) = inverse document frequency — log[(N − n(q_i) + 0.5) / (n(q_i) + 0.5) + 1], where N = total documents, n(q_i) = documents containing q_i. Rare terms receive higher weight.
k₁ = term-frequency saturation parameter (typically 1.2–2.0). Controls how quickly repeated occurrences of a term stop increasing the score.
b = length-normalization parameter (typically 0.75). At b=1, full length normalization; at b=0, no length normalization.
avgdl = average document length across the collection

Under the hood: Vector Indexing (ChromaDB)

What it does: ChromaDB is the vector database used internally by Smart Search and Semantic modes. When you upload documents, every passage is converted into a row of numbers (a vector) and stored in a collection. ChromaDB indexes those rows so that when you search, it can find the passages whose rows are nearest to your query’s row without re-reading every passage from scratch.

Why it matters: Comparing your query against every passage one row at a time gets slow as an archive grows. A vector database keeps the rows in a structure built for nearest-neighbour lookup — like organizing a library by subject so you can go straight to the right shelf instead of walking every aisle. This keeps semantic search fast even as your collections grow.

Mathematical notation

Nearest-neighbour retrieval over a dense vector space. Text passages are encoded into fixed-dimensional vectors and stored in a vector database (ChromaDB), which returns the indexed vectors closest to the query vector under a chosen distance function. NN(q) = argmin_{v ∈ V} d(q, v)

q = query vector (the embedding of the search query)
V = the set of all indexed passage vectors in the matrix
v = an individual passage vector (one row of the matrix)
d(q, v) = a distance function between vectors, e.g. cosine distance, inner-product (dot product) d = −∑ q_i · v_i, or L2 (Euclidean) distance d = √∑(q_i − v_i)²
NN(q) = the nearest neighbor — the passage vector closest to the query

Under the hood: Semantic Embeddings

What it does: The configured embedding model (part of the Sentence-BERT family of transformer encoders) reads your collections and decides what each passage means. When you uploaded your documents, the model read each passage and produced the row of numbers (the vector) that represents its meaning in the matrix. It does the same for your search query. It understands meaning because it was trained on enormous amounts of text and learned that, for instance, “workers’ revolt” and “proletarian uprising” refer to the same idea. The search then compares your query’s row of numbers against each passage’s row: passages whose rows closely match yours appear at the top of your results, because close rows mean close meaning.

When it activates: Semantic embeddings power the Smart Search and Semantic modes. When you search for “economic crisis,” the embedding model also surfaces passages about “financial panic,” “commercial collapse,” or “overproduction” from your collections, even if those exact words weren’t in your query.

Why it works: Traditional keyword search treats “dog” and “hound” as completely unrelated because they share no letters. The embedding model learned from billions of sentences that these words appear in similar contexts and mean similar things, so it places them in nearby rows of the matrix. It compares meaning by measuring how closely two rows point in the same direction in the high-dimensional space of the matrix — if two passages discuss the same topic, their rows will point the same way regardless of the specific words each one uses.

Mathematical notation

A modification of the BERT transformer that produces semantically meaningful sentence embeddings using siamese and triplet network structures. Passages and queries are independently encoded, then compared via cosine similarity. sim(q, p) = cos(θ) = (q · p) ⁄ (||q|| · ||p||) = ∑_i=1^d q_i p_i ⁄ (√∑ q_i² · √∑ p_i²)

q = query embedding vector, produced by passing the query through the BERT encoder followed by mean pooling
p = passage embedding vector, produced identically from the indexed passage
d = embedding dimensionality (typically 384 or 768 depending on model)
q_i, p_i = the i-th component of each vector
||q||, ||p|| = Euclidean norms (magnitudes) of each vector
cos(θ) = cosine of the angle between vectors. Values range from −1 (opposite meaning) to +1 (identical meaning). By normalizing for magnitude, cosine similarity captures semantic direction only.

Smart Search (Hybrid)

What it does: Smart Search runs two searches over your collections at once — a keyword search (BM25) and a meaning-based matrix search (vector) — then intelligently merges their results into a single list. It works like asking two research assistants with different strengths to each find relevant passages, then combining their recommendations. A passage that both assistants independently flag as important rises to the top of your results; a passage found by only one still appears, but further down the list.

When to use it: This is the best default for most research. Use it when you want to be thorough — when you want to find passages that contain your exact words and passages that discuss the same idea in different language. It protects you from the blind spots of either method used alone: keyword search might miss relevant synonyms, and meaning-based search might miss rare proper names. Smart Search catches both.

Why it works: Each search method has strengths the other lacks. Keyword search is precise and predictable — if a word exists anywhere in your collections, it will find it. Meaning-based search understands synonyms and paraphrasing but can sometimes drift from what you typed. By combining both and promoting passages that appear in both result lists to higher positions, Smart Search gives you the reliability of exact matching together with the breadth of semantic understanding.

Mathematical notation

Combines BM25 keyword scoring with vector similarity using Reciprocal Rank Fusion (RRF) to merge the two ranked lists into a single result set. RRF(d) = ∑_{r ∈ R} 1 ⁄ (k + rank_r(d))

d = a document appearing in one or more result lists
R = the set of ranking systems (here: BM25 and vector search)
rank_r(d) = the position of document d in ranking r (1 = top result)
k = a smoothing constant (typically 60). Prevents top-ranked documents from dominating; higher values give more weight to documents that appear in multiple lists.

Scope & Archives

Searches cover your whole library across all collections. To narrow a search, open Advanced Search: date range, author, document type, collection type, and structured AND/OR/NOT rows all apply to every view — Concordance, Dispersion, Frequency, and Relevance — and active filters appear as removable chips above the results.

This setting only affects searching here in the Reading Room. Choosing which collections or documents a piece of research uses is done separately, in the Atelier.

Search Settings

Passage Size: This is set in Resource Management, not here. It controls how long each prepared passage is, measured in tokens (roughly, word-pieces), and you can choose 256, 512, 1,024, or 2,048. Smaller passages make searches more precise and give tighter quotations, but they hold less of the surrounding context; larger passages keep more context but can make matches less focused.

Advanced Search

Structured Query: Each row of a structured query lets you pair a field — the title, the author, the full text, or all fields — with a term to look for. When you add more than one row, Mira combines them using the AND, OR, or NOT operator you choose between them. You can use structured rows on their own, even with the main search box empty — for example, just Author = Marx to find everything by that author. (How the Relevance view ranks results is chosen with the “Rank by” picker under the view tabs, described above, not here.)
Boolean Operators: These let you combine conditions. AND requires both conditions to be met, OR matches a result that meets either one, and NOT excludes any result that contains that term.
Date Range: You can restrict results to a range of publication years by filling in a “from” year, a “to” year, or both. This applies to all four views.
Document Type: You can restrict results to particular kinds of document — for example an article, a book, or a manuscript — as detected when the file was prepared. This applies to all four views.
What is not supported yet: A few advanced query features are planned but not yet available: proximity operators (finding two words within a set distance of each other), wildcards (such as “colon*” to match anything starting with those letters), quoted exact-phrase searches, and lemma matching (so that “colonise” would also match “colonising”). For now, every match is a literal piece of text.

Atelier

What the Atelier is

The Atelier is where you ask a research question and let a small team of helpers go looking for answers inside the PDFs you have added to Mira. We call each helper a “subagent”: it is one language-model worker that takes on a single slice of your question. You type your question in the box at the bottom; the Atelier breaks the question into several lines of inquiry, sends one subagent down each line, and each subagent reads through your documents and gathers the passages that bear on the question. When they are finished, their findings are woven together into a single essay that answers you, and every point in that essay is tied back to the exact passage it came from.

Asking a question

The bar at the bottom of the Atelier has just three parts.

The text box is where you type your question. Specific, focused questions give far better results than broad ones.
The flask button to the left of the box, labelled “Multi-Agent Architecture,” opens the designer where you decide how the work is carried out — how many subagents to send out, which models they use, what they are allowed to read, and more. A separate, plain-language guide to every control in that designer can be opened from the main Settings button.
The round button to the right of the box is a single button that changes as the work proceeds. Click it to send your question; while the subagents are working it becomes a spinner you can click to stop the run; once a run has finished or been stopped it becomes a “try again” arrow so you can run the same question afresh. There is no separate stop button.

Setting up a run before you send it

Open the “Multi-Agent Architecture” designer (the flask button) to decide what the subagents may look at and how they work. The controls you will reach for most often are these.

Two scope buttons let you choose what the subagents may read. “All sources in library” lets them read everything you have added to Mira, while “Selected sources only” limits them to the documents you have picked. Use “Selected sources only” when you want a run kept tightly focused on a few texts.
The “Context” button opens a small editor where you can leave the subagents background notes about your project, so they understand what you are really after. A dot on the button shows when you have notes saved.
The switch labelled “Also include my active Writing Desk document as context” hands whatever you currently have open in the Writing Desk to the subagents as further background for the run.

Quote cards, and why they are always exact

As the subagents find passages worth quoting, the Atelier shows each one as a quote card. A quote card is a short slab of text lifted straight from one of your PDFs, shown above its citation. The wording on a card is always the original source, word for word — Mira never paraphrases it and never lets a subagent rewrite it. Each card shows three things.

It shows the quoted passage itself.
It shows a citation in Chicago style, including the page number where the passage appears.
It shows a small coloured dot that reports how the quote was checked against the source. A “verified (exact)” dot means the words were found in the document exactly as shown, while the other states warn you that the match was only close, or that the passage could not be found and should not be trusted.

When you hover over a card, a row of small buttons appears: you can open the source in the reader, show the surrounding context in a popup, copy the citation, or use “Send to writing desk,” which drops the exact passage and a matching footnote straight into whatever you are writing. A citation is only ever as trustworthy as the document behind it, so always check a passage against the original — see the warning about pieced-together PDFs in the Resource Management guide.

Your margin notes as context

If you have written margin notes on your documents in the Reading Room, the switch labelled “Include my annotations as context” in the Multi-Agent Architecture designer decides whether the subagents get to see them. When it is on, your notes are handed to the subagents as background and are weighted heavily when deciding which passages are relevant to your question, so a note can pull a passage you cared about to the front of the search. Your notes are kept strictly separate from the source text, however: a note is never quoted and can never end up on a quote card. Quote cards only ever contain the original words of the document. In short, your notes guide the search but never become evidence.

Watching a run unfold

The subagent grid shows one live tile for each subagent, so you can watch them all work at the same time. “Expand all” opens every tile at once.
The reasoning trace is a section you can open to read the subagents' step-by-step thinking as it streams in.
The “Loaded Sources” panel lists the documents that have been pulled into the run, and “Clear” empties that list.

The run record

Every run is saved as a permanent record you can reopen later, and you can download it as a JSON file. JSON is simply a plain-text file that stores information in a tidy, labelled form that both you and other programs can read. The record holds the full question and instructions, each subagent's complete transcript (what it was told, how it reasoned, and what it answered), the passages it retrieved, and a line-by-line account of every request that was made — so you can check exactly what each subagent was given and how it replied.

Settings

The gear icon in the Atelier opens “Atelier Settings,” which controls how this view looks and behaves: the text size, whether the empty composer suggests research questions drawn from your library, and the reveal animations. It does not configure the research itself.

Everything about how a run is actually carried out — how many subagents work at once, which models they use, how deeply they search, what they may read, and the cost limits — lives in the “Multi-Agent Architecture” designer (the flask button beside the question box). A separate, detailed guide explains every one of those controls in plain language, and you can open it from the main Settings button.

Tip: the more specific and focused your question, the better the result.

Writing Desk

The Writing Desk is a word processor built into Mira for writing notes, drafts, and analysis next to your documents. It reads and writes Word files (the .docx format). Mira provides the file buttons along the top of the page; the formatting buttons just above the page itself — bold, italics, headings, lists, and so on — come from the built-in editor, SuperDoc.

Opening and saving your work

By default the Writing Desk works like any website: “Open” uploads a Word file from your computer, and “Save” or “Download” hands you a fresh copy to keep, with no permissions ever requested. If you would rather edit a file on your computer in place, turn on “Save directly to disk” in the Writing Desk settings (the gear icon); your browser will then ask permission once, and afterwards your edits are written straight back to that file. This in-place editing needs Chrome, Edge, or Brave — Safari and Firefox do not offer websites that ability, so there Mira saves with Download instead. “Save As” lets you choose a new name or location.

Your work is saved automatically. Every few seconds a copy is tucked away inside your browser, so even if you reload the page or decline a permission prompt, nothing is lost. When a file on your computer is linked, your edits are also written back to it about every thirty seconds.

The file buttons

“Open” loads a Word file to edit, and “Download” saves a copy to your computer at any time.
“Save” and “Save As” store your work; when “Save directly to disk” is off, they simply hand you a downloaded copy.
“Reload from disk” re-reads the linked file from your computer, and “Unlink” detaches it so that saving goes back to giving you copies. These two buttons only appear when “Save directly to disk” is turned on in a browser that supports it.
“Find” searches your document, “Print” sends it to your printer, and “Comments & tracked changes” opens a panel for margin comments and a record of edits.

Projects, drafts, and sections

Your work is organised in three simple levels. A project (such as “Dissertation”) holds your drafts, and each draft holds its sections. A draft is one piece of writing — it might be a chapter, an article, or a set of research notes, whatever you are working on — and you give it whatever title you like (for example, “Beyond the Pale”). A section is one document inside that draft. In the list on the left you see each draft as a heading with its sections nested underneath, numbered for you: draft 3, then 3.1, 3.2, and so on.

Open the “Manage” window to build this. It works like a tidy file browser: on the left is a tree of your project, drafts, and sections; on the right are the contents of whatever you click, which you can show as a list or as cards. Press New — it offers New chapter / draft (a container that holds documents) and New document (a single Word file, placed inside the chapter you are looking at). Each choice says in plain words what it makes, so you won’t create a document when you meant a chapter. You type the title; Mira keeps the numbers in order.

You can drag and drop to organise: drag a document onto a chapter to move it inside, drag it onto the project name to take it back out, and drag a chapter onto another chapter to reorder them. Dragging one document above or below another in the same chapter reorders them. You can also rename or delete a chapter or document, or use a document’s “Move to…” button instead of dragging. Keeping each document to roughly thirty-five pages or less makes it open and type quickly — so a long chapter is best kept as a few numbered documents.

Sharing with other people

Every project, chapter, and document has a Share button in the Manage window. Click it and type someone’s email to give them access. You choose what they can do: Viewer (read only), Commenter (read and comment), Editor (read and edit the writing), or Manager (edit and also reorganise — add, rename, move, reorder, and delete chapters and documents). Only the owner (you) can delete the whole project or hand it to someone else. If the person doesn’t have a Mira account yet, Mira emails them an invitation. Sharing a whole project or chapter shares everything inside it; you can also share a single document on its own. You can change someone’s access or remove them at any time, or make a link that anyone you give it to can open.

Documents other people have shared with you appear under Shared with me. Live co-editing is automatic: whenever two or more people have the same shared document open, everyone’s changes appear in real time, and edits save back to the owner’s copy — there is no setting to switch on. When you’re the only one in a document it behaves normally; if someone edits it outside a live session and saves first, Mira warns you so your work is never quietly overwritten.

To hand a project to someone else permanently, open the project’s Share dialog and choose Transfer ownership. The whole project moves to them and they become the owner; you keep access as a Manager. Only the current owner can do this.

Bringing quotations in from the Atelier

On any Atelier quote card, the “Send to writing desk” button drops that exact passage into your draft at the cursor, together with a matching Chicago-style footnote. The wording is always the original source, word for word, and the footnotes are numbered for you as you add them.

Architecture Designer Guide

This designer is where you decide how a research run is carried out. There is no separate run button: whatever you set here is applied straight away, and the run begins when you type a question in the Atelier and press send. A “subagent” is one language-model worker that takes on a single slice of your question, and a run usually sends several of them out at once.

How a run works

Your question travels through a short pipeline, shown as the row of stages along the top of the designer.

The planner reads your question first and decides which angles are worth pursuing and how many subagents to send out.
The subagents then work at the same time, each one searching your documents, reading the passages it finds, and writing up what it has learned. “Investigation depth” sets how hard each one digs, and “Concurrency” sets how many work at once. With “Shared workspace” on, they can build on each other’s findings.
Aggregation gathers everything they found. With “Dedup” on, passages that several subagents turned up are merged so the same source is not repeated.
Debate is an optional extra round in which the subagents criticise each other’s findings to catch mistakes before the essay is written.
Finally, the essay stage writes your answer from everything that was gathered.

At present the run uses the “Exhaustive audit,” which is on by default. Instead of a single pass it sends wave after wave of subagents, drilling into the most important and most disputed points until there is nothing new to find or it reaches the limits you set, and then writes an article-length answer. The Direct and Dialectical strategy choices shown beside the essay stage are turned off for now and will return in a later version.

What the subagents may read

The “Sources” section sets the reading list. “All sources in library” lets the subagents read everything you have added to Mira, while “Selected sources only” limits them to the documents you pick in the panel just below. “Web research” adds a subagent that looks up outside scholarship from two free scholarly catalogues, OpenAlex and Crossref, kept clearly separate from your own documents; it needs at least three subagents to take effect.

Keeping the answer careful and affordable

The “Reasoning depth & rigour” controls make the run weigh its evidence carefully: it keeps a tally of each claim alongside the evidence for and against it, and it forbids conclusions that sound more certain than the evidence allows. The spending figures under “Compute & cost limits” (“Target $/agent” and “Target $/run”) are shown for your awareness only and do not stop a run. The “Agent cap” and “Synth cap” limit how much each subagent and the final essay writer may produce; if one reaches its cap, its work is trimmed back to the last complete sentence rather than lost.

Your margin notes

“Include my annotations as context,” in the “Retrieval & intake” section, decides whether your Reading Room margin notes are handed to the subagents. When it is on, your notes are used as background and weighted heavily when finding relevant passages, but they are kept strictly apart from the source text and can never be quoted on a card. Your notes steer the search; they never become a quotation.

Saving, sharing, and the gear

“Save as preset” stores your current set-up under a name so you can return to it; “Export” and “Import” move a set-up to or from a file you can back up or share; “Logs” opens the full record of the run currently loaded in the Atelier; and “Done” simply closes the designer, since your set-up is already applied. The gear icon at the top opens “Atelier Settings,” which only changes how this view looks and behaves — it does not configure the research itself.

For a control-by-control explanation of every option in this designer, open the main Settings button and choose the “Multi-Agent Guide” tab.

Reading Room

Collection type

Year range

Author

Document type

Begin Your Research

Main Library

Resource Management

All Clear

Special Collections

No Collections Yet

Add Content

Public Domain Libraries

Internet Archive — Texts

Standard Ebooks

Perseus Digital Library

Upload Your Own

Upload Files

Acquisitions

Atelier

Library Metadata Review

Writing Desk

Session Logs

Process Observation

Reading Room

Collection type

Year range

Author

Document type

Begin Your Research

Main Library

Resource Management

All Clear

Special Collections

No Collections Yet

Add Content

Public Domain Libraries

Internet Archive — Texts

Standard Ebooks

Perseus Digital Library

Upload Your Own

Upload Files

Acquisitions

Atelier

Library Metadata Review

Writing Desk

Session Logs

Process Observation

Multi-Agent Architecture

Sources & materials

Scale & depth

Rigor

Token caps & retrieval

Delivery

Agents

Save Preset

Privacy Doctrine

Atelier Settings

Research context

Library Settings

Acquisitions

What this tab is for

Try this first: add a book from an online library

The two tabs

The online libraries you can search

The buttons on each result (they appear after you search)

Advanced Search — search by title, author, or year

Adding your own PDFs (the Local Library tab)

Can't find your library?

If you see a "Bridges" tab

Resource Management

What is this tab?

Inventory tab

What every pill means (in plain words)

How your files become searchable

Metadata tab — checking titles & authors

Always check the citation

Vector Inspector (coming soon)

The settings gear (top corner)

Main Library

What is the Main Library?

The most common thing you’ll do here

What each item shows

Right-click an item — every choice explained

Working with several items at once

How files get prepared

The settings gear (top corner)

Special Collections & Archives (coming soon)

Reading Room

Search Views — choose the question you’re asking

How your collections are searched

Retrieval Methods

Scope & Archives

Search Settings

Advanced Search

Atelier

What the Atelier is

Asking a question

Setting up a run before you send it

Quote cards, and why they are always exact

Your margin notes as context

Watching a run unfold

The run record

Settings