LLM Chat with Context

chat_with_context() assembles a token-budgeted context window from the project graph and sends it to an LLM. This vignette shows real model responses captured from a live run — no API key is required to read the vignette.

How it works

library(rrlmgraph)

graph <- build_rrlm_graph(system.file("extdata", "demo", package = "rrlmgraph"))

# GitHub Models Marketplace — free, uses GITHUB_PAT in Actions
answer <- chat_with_context(
    graph,
    "How does the data preparation and validation pipeline work?",
    provider = "github",
    model    = "gpt-4o-mini"
)
cat(answer)

Under the hood chat_with_context():

1. Calls query_context() to run relevance-guided BFS capped at budget_tokens (default 2 000).
2. Assembles a system prompt from the selected node summaries and source snippets.
3. Sends the prompt + user query to the chosen provider via the ellmer package.
4. Appends an entry to .rrlmgraph/task_trace.jsonl for future weight updates.

Captured outputs

The responses below were generated once by the precompute-vignettes GitHub Actions workflow and committed to the repository as vignettes/precomputed/llm_chat_outputs.rds. They will not be re-generated on every R CMD check or CRAN build.

# When a vignette is built, cwd = vignettes/ so this relative path works
# both during R CMD check and pkgdown site builds.
precomputed_path <- file.path("precomputed", "llm_chat_outputs.rds")
if (file.exists(precomputed_path)) {
    outputs <- readRDS(precomputed_path)
} else {
    outputs <- NULL
}

if (is.null(outputs)) {
    cat(
        "> **Note:** Precomputed outputs not yet available.",
        "Run the `precompute-vignettes` workflow to generate them.\n"
    )
} else {
    for (item in outputs) {
        cat("### Query\n\n")
        cat(item$query, "\n\n")
        cat("**Context nodes used:**", paste(item$context_nodes, collapse = ", "), "\n\n")
        cat("**Provider / model:**", item$provider, "/", item$model, "\n\n")
        cat("**Response:**\n\n")
        cat(item$response, "\n\n")
        cat("---\n\n")
    }
}

Note: Precomputed outputs not yet available. Run the precompute-vignettes workflow to generate them.

Regenerating outputs

Trigger the precompute-vignettes workflow from GitHub Actions to refresh llm_chat_outputs.rds and redeploy the pkgdown site.