Skip to contents

Chat with an LLM using graph-derived context

Source: R/llm_interface.R
chat_with_context.Rd

Retrieves a token-budgeted context window from the graph via query_context(), builds a grounded system prompt, and sends a message to an LLM. Requires ellmer, which handles provider connections and the tool-call loop that implements RLM-Graph traversal.

Usage

chat_with_context(
  graph,
  message,
  provider = c("openai", "ollama", "github", "anthropic"),
  model = NULL,
  budget_tokens = 6000L,
  seed_node = NULL,
  min_relevance = 0.1,
  ...
)

Arguments

graph

An rrlm_graph / igraph object.

message

Character(1). User message / question.

provider

Character(1). LLM provider. One of "openai" (default), "ollama", "github", "anthropic".

model

Character(1) or NULL. Model identifier. When NULL (default) a sensible per-provider default is used: "gpt-4o-mini" (openai/github), "llama3.2" (ollama), "claude-3-5-haiku-latest" (anthropic).

budget_tokens

Integer(1). Context token budget passed to query_context(). Default 6000L.

seed_node

Character(1) or NULL. Forwarded to query_context().

min_relevance

Numeric(1). Forwarded to query_context().

...

Additional arguments forwarded to the ellmer chat_*() constructor (e.g. base_url for ollama).

Value

Character(1) containing the LLM response text. Throws on error.

System prompt structure

The system prompt combines:

  1. The assembled context string (code snippets relevant to the query).

  2. A grounding constraint block that instructs the model to answer only from the provided context and to cite node names.

Authentication

"openai"

Set OPENAI_API_KEY environment variable.

"ollama"

No key required (local daemon). Set OLLAMA_BASE_URL to override the default http://localhost:11434.

"github"

Set GITHUB_PAT environment variable. Requires access to the GitHub Models Marketplace (https://github.com/marketplace/models).

"anthropic"

Set ANTHROPIC_API_KEY environment variable.

See also

query_context(), update_task_weights()

Examples

if (FALSE) { # \dontrun{
g <- build_rrlm_graph("mypkg")
# OpenAI (default) -- requires OPENAI_API_KEY
chat_with_context(g, "How does data_prep() work?")
# GitHub Models Marketplace -- requires GITHUB_PAT
chat_with_context(g, "How does data_prep() work?", provider = "github")
# Local Ollama
chat_with_context(g, "How does data_prep() work?",
  provider = "ollama", model = "llama3.2")
} # }