How ragX works

ragX answers in three stages: a forage loop, a slot chain and a single compose call. The first two are separate routes to the same goal, a query that names evidence the question does not name. The compose call answers over everything both routes found.

The problem it solves

Single-shot RAG retrieves with the question as the query. A question such as

Who is the spouse of the actor who plays Jarvis in Iron Man?

names Iron Man and Jarvis, but not Paul Bettany, whose article holds the answer. On all of Wikipedia that answer passage ranks 146,602th for the question itself, so no query copied from the question reaches it. As a corpus grows, single-shot retrieval loses such unnamed evidence about four times faster than named evidence, and neither a larger embedder nor a deep reranker changes that. A query that names the missing entity does.

1. The forage loop

An agent explores the corpus with two tools, one action per turn:

Action	What it does
RANK: <keywords>	ranks every passage with the variant's ranker and lists the top 8 by name and title
READ: <passage>	prints one passage in full, with line numbers
ANSWER: <short answer>	ends the loop with a draft answer

The loop runs for at most ten turns. Because the agent writes each query after reading, a later RANK can name an entity it found in an earlier READ: in the example, reading the Iron Man article leads to RANK: Paul Bettany spouse.

Every observation is fenced as untrusted data before the agent sees it.

2. The slot chain

Separately, the question is decomposed into at most four ordered sub-questions, with #N slots for earlier answers:

Who plays Jarvis in Iron Man?
Who is the spouse of #1?

Each hop is then resolved in turn: the sub-question, with the earlier answers filled into its slots, retrieves four passages; one short call answers it from their most relevant sentences; and that answer fills the next slot. The two top passages of each hop are kept for the compose step, because a hop's passage often shares no words with the original question.

The ragx-fast variant skips this stage.

3. Compose

One call answers the question over, in order:

1. up to 8 passages the agent read,
2. up to 12 extra passages: first those that resolved a hop, then the other passages the agent surfaced, ranked against the question,
3. in ragx and ragx-think only, the backfill: the question's own top 48 passages that are not already included.

Passages are composed whole. If the evidence would exceed the model's context window, ragX trims from the end, removing backfill passages first. The compose prompt asks for the shortest exact answer and adds a one-line hint about the expected answer type (a date, a place, a person, yes or no, one of the named options).

The answer is then grounded: it is cited to the composed passages that contain it.

Calls per question

A run makes one call per forager turn, one decomposition call, one call per hop and one compose call, so typically 8 to 15 calls. ragx-fast makes only the forager calls and the compose call. See Cost and speed.

The frozen settings

The defaults are the settings of the paper, and every variant shares them:

Setting	Default	Meaning
steps	10	forager turns
rank_k	8	passages listed per RANK
hop_k	4	passages retrieved per hop
max_hops	4	sub-questions at most
read_cap	8	read passages composed
extra_cap	12	extra passages composed
backfill	48 or 0	the question's own top passages appended (48 in ragx and ragx-think)
tools	RANK, READ	forager tools (SEARCH can be added)
fence	on	tool output fenced as untrusted data

All of them can be changed; see Configuration.