Settings → Retrieval

This page changes no setting. It is a window on what memory retrieval actually did: every search the chat runs against your vaults is written to a journal, with its scope, its tuning, its timings and the sources it brought back. The rail entry is Retrieval; the page is titled Retrieval journal and its subtitle reads Every memory search records its scope, tuning (top-k / mode / rerank / recency), timings and the sources it retrieved with their scores. A red top-score flags a weak retrieval — a cue to raise top-k, switch query mode, or relax the rerank.

For how retrieval works, see Memory recall. The knobs the subtitle mentions (top-k, mode, rerank) are set in the chat's memory controls, not here.

Everywhere on this page, a score is coloured the same way: green from 0.60 up, amber from 0.35, red below, and a grey — when there is no score at all.

Tuning comparator

A collapsible card, closed by default, headed Tuning comparator, Run one query with two settings and compare. Opening it loads your vault list once.

• Query: Query to test…. Enter runs the comparison.
• Scope: all vaults (default, a federated search) or one vault by name.
• Compare runs both sides; the button reads Searching… meanwhile. An empty query answers Enter a query first.

Two panes, A and B, each with the same four controls. The pane names and the three captions are shown in English in every language:

Control (as shown)	Values	Default A / B
top-k	1 to 50	5 / 10
mode	Default (sends nothing), HYBRID, UNIFIED	Default
rerank λ	0 to 1, step 0.1; empty is not sent	empty
recency-biased (checkbox)	on / off	off

Each pane then shows its top score (coloured), a trophy on the winner (the higher top score; A wins a tie, and the winning pane gets a green border), the number of sources and the time spent (embedding plus search, in ms), then every source: score, vault name, a two-line preview. No sources retrieved. means that side came back empty. A side that failed prints its error, or "A failed" / "B failed" in English when there is none.

1. Open the card

   Click Tuning comparator. The vault list loads.

2. Type the query and pick a scope

   Use a question you know the answer to, and the vault it should come from, or all vaults.

3. Set the two sides

   Leave A as it is and change one thing on B: a higher top-k, another mode, a rerank λ, or recency-biased.

4. Compare

   Read the two top scores and the sources under them. The trophy only says which top score is higher; the sources tell you whether the right memory came up.

Retrieval only, and it is journaled

The comparator retrieves and never generates: no model writes an answer. But each Compare embeds the query twice and searches your vault(s), so the two searches appear in the journal below like any other.

Journal toolbar

• Refresh reloads the last 200 entries.
• Clear empties the journal, after a native confirmation: Clear the whole retrieval journal? This only removes the local log, not your memory. It is greyed while the page is busy or the journal is already empty.
• On the right, the count: {count} retrievals.

The journal

While it loads: Loading the journal…. If it cannot be read: a red box with the error, or Could not load the retrieval journal. With nothing recorded yet: No retrievals yet. Chat with memory enabled (a vault scope) to populate the journal.

Each entry is a collapsible row: a chevron, the top-score chip (hover: Best similarity score; — when the search returned nothing), the query (∅ if it was empty), the scope (all vaults or the vault's name, — if unknown), and the number of sources.

Expanded, the entry shows a technical line, in English: the timestamp, top-k=N, mode=… and λ=… when they were set, recency-biased when the recency bias was on, embed=Nms and search=Nms. Then the sources, best score first: score, vault name, spine type (raw id), file name with a document icon (hover: the full path), and a three-line preview; or No sources retrieved.

Reading a red chip

A red top score means the best memory found was a weak match. Before changing anything, open the entry: if the right file is there with a low score, the query was phrased far from the memory; if it is absent, the vault scope was probably too narrow, or the memory is not in a vault yet.