How retrieval works

Onadit looks for a passage 2 ways at once, by keyword and by meaning, puts both on 1 scale, and nudges recent work up. This page gives the formula.

2 lanes#

Every search runs 2 lookups at the same time and merges the results by ID.

Lane	How	Needs
Keyword	A full-text index over the archive, scored with BM25	Nothing. Always on
Meaning	Your question is embedded, then compared with stored embeddings by cosine similarity	Search by meaning turned on

Each lane fetches 3 times the number of results you asked for, and at least 30. A row found by both is marked hybrid.

With Search by meaning off, only the keyword lane runs. The response says "mode": "lexical".

The score#

text

score = 0.55 × similarity
      + 0.10 × recency
      + 0.10 × quality
      + 0.10 × tier
      + 0.10 × 1.0          (reserved for recorded outcomes, fixed today)
      + 0.05 × found-both-ways
      + 0.05 × freshness

Term	Value
Similarity	For a meaning match, the cosine similarity. For a keyword match, its BM25 score divided by the best keyword score, then scaled to the best meaning score in the same pool. For a row found both ways, the larger of the 2
Recency	`1 / (1 + 0.01 × days)`. 1.0 today, 0.77 after 30 days, 0.5 after 100
Freshness	`exp(-days / 14)`. 1.0 today, 0.61 after 1 week, 0.37 after 2 weeks, 0.12 after 1 month
Quality	1.0, 0.7 or 0.4 for conversations rated high, medium or low. 1.0 for documents and digests
Tier	1.0 for current conversations and documents, 1.1 for digests, 0.7 for archived conversations
Found both ways	1.3 if both lanes found the row, otherwise 1.0

days counts from the time a row was indexed. For an imported conversation that is the day of the import.

Similarity carries 55% of the score. The 2 time terms together carry 15%, so a recent passage wins a close call and loses to a clearly better match.

After ranking#

Near-identical copies of the same conversation are folded into 1 result.
Generated reports are placed behind first-hand material.
Each result is cut down to a snippet (400 characters) and a summary (800), with the size of the full text and a pointer to open it.

Current and archived conversations#

New conversations land in the current table. rift compact moves those older than 30 days to an archive table and, if you have enrichment on, writes a weekly digest of them. Compaction runs when you call it. It is not scheduled.

A normal search covers documents, current conversations and digests. The archive is searched too in 3 cases:

Search by meaning is off.
The best result scores under 0.4.
An agent asks for it with rift_conversations_search and detail: "full".

When a search is degraded#

"degraded": true means the meaning lane was tried and did not finish, so the results are keyword only. The response gives the reason. The usual one is that embedding the question took longer than 1.5 seconds, the most Onadit will wait.

Search by meaning being off is not a degraded search. It is the keyword mode working as intended.

How a context pack is built#

rift_context_pack runs the same search, then sorts what it finds for an agent that is about to work.

First it reads the task and picks an intent:

Intent	When	Effect
`current_truth`	The task asks what is true now	Live files and trackers come first
`reasoning_archive`	The task asks why something was decided	Older conversations are treated as valid evidence
`blended`	Both, or neither	A mix

Then it gives every source a level of trust:

Level	What it is	Examples
`live_state`	Things that can be run or observed	Source code, install scripts, `package.json`, `config.json`
`tracker`	Files kept to record state	`PROJECT_STATE.md`, `TODO.md`, dated decision notes
`committed_doc`	Durable written framing	`README`, `CLAUDE.md`, `AGENTS.md`, a PRD
`discussion`	Conversations and agent reasoning	Any captured or imported chat
`discounted`	Material that may be stale	Untracked or deleted files, generated reports

A conversation is never the only authority for what is true now. When the pack has no live_state item to stand on, it says so: "current_truth_caveat": "tracker-backed, not live-verified". When it knows which files would settle the question, it lists them and the agent should read them.

A decision or an example needs a relevance of at least 0.3 to be included. Items are dropped from the end until the whole response fits in max_bytes.

The embedding model#

With a key Onadit provides	`voyage-4-lite`, 512 dimensions, served by MongoDB Atlas
With your own Voyage key	`voyage-3-lite` by default, 512 dimensions. Change it with `embedding.model`
Sources set to `"extraction": "local"`	An Ollama model on your Mac, 768 dimensions
With no key	No embedding. Rows are stored without a vector and found by keyword

Embeddings are cached on disk, so the same text is not sent twice.

Last updated 12 October 2026. This page is also available as Markdown, and every page is listed in llms.txt.