Query the rrlm graph and retrieve a token-budgeted context

Performs a relevance-guided breadth-first search starting from a seed node and builds a context string that fits within budget_tokens. The budget is a hard constraint: the function never returns more tokens than requested.

Usage

query_context(
  graph,
  query,
  seed_node = NULL,
  budget_tokens = 2000L,
  min_relevance = 0.1,
  max_nodes = 20L,
  method = NULL,
  verbose = FALSE
)

Arguments

graph

An rrlm_graph / igraph object created by build_rrlm_graph().

query

Character(1). User query string.

seed_node

Character(1) or NULL. Name of the vertex to start traversal from. NULL (default) triggers automatic selection: the function-type node with the highest PageRank.

budget_tokens

Integer(1). Hard token limit. Default 2000L.

min_relevance

Numeric(1). Minimum relevance score \([0, 1]\) for a node to be admitted. Default 0.1.

max_nodes

Integer(1). Maximum number of nodes (excluding the seed) to absorb. Default 20L.

method

Character(1) or NULL. Embedding method passed to embed_query(). NULL (default) reads the "embed_model" graph attribute.

verbose

Logical(1). Print progress messages via cli_inform(). Default FALSE.

Value

A named list with class c("rrlm_context", "list"):

nodes

Character vector of absorbed node names, relevance-ordered, seed first.

context_string

Character(1) assembled by assemble_context_string().

tokens_used

Integer(1). Approximate token count of context_string.

budget_tokens

Integer(1). The budget that was used.

seed_node

Character(1). The seed node name.

relevance_scores

Named numeric vector of relevance scores for every absorbed node (seed included).

Algorithm

1. Embed query with embed_query().

2. Identify the seed node: if seed_node is NULL, select the function-type vertex with the highest pre-computed PageRank; otherwise validate and use the supplied name.

3. Initialise visited = {seed} and frontier = neighbours(seed).

4. BFS loop while tokens_used < budget_tokens and frontier is non-empty:

   • Score every frontier node with compute_relevance().

   • Select the best-scoring node with score >= min_relevance.

   • Compute its token cost (.count_tokens()).

   • Accept the node only if adding it stays within the budget; otherwise skip and try the next-best.

   • Mark as visited; expand its neighbours into the frontier.

5. Call update_task_weights() to update the learning trace.

6. Assemble the final context string with assemble_context_string().

See also

compute_relevance(), build_node_context(), assemble_context_string(), update_task_weights()

Examples

if (FALSE) { # \dontrun{
g   <- build_rrlm_graph("mypkg")
ctx <- query_context(g, "load training data", budget_tokens = 1000L)
cat(ctx$context_string)
} # }