rlm

name: rlm description: Recursive Language Model pattern for analyzing data too large for a single context window

RLM (Recursive Language Model) Skill

When to Use

Use this pattern when:

• Data exceeds ~50KB (logs, codebases, datasets)
• Single-pass analysis produces low quality results
• You need to aggregate insights from multiple data segments

Available Functions (in iex)

• llm_query.(prompt, data) — Query sub-LLM with default model
• llm_query_with_model.(model, prompt, data) — Query with specific model
• llm_query_many.(chunks, prompt) — Parallel batch query (default concurrency: 4)
• llm_query_many.(chunks, prompt, max_concurrency: 8, model: "gpt-4o") — With options

Pattern: Explore → Chunk → Analyze → Synthesize

Step 1: Load & Explore

iex> raw = File.read!("/path/to/large_file.log")
iex> lines = String.split(raw, "\n")
iex> length(lines)  # Check size
iex> Enum.take(lines, 5)  # Sample first few lines

Step 2: Chunk

iex> chunks = Enum.chunk_every(lines, 500) |> Enum.map(&Enum.join(&1, "\n"))
iex> length(chunks)  # How many chunks?

Step 3: Parallel Analyze

iex> results = llm_query_many.(chunks, "Identify error patterns and anomalies in this log segment. Return a bullet list.")

Step 4: Synthesize

iex> combined = Enum.with_index(results, 1) |> Enum.map(fn {r, i} -> "## Chunk #{i}\n#{r}" end) |> Enum.join("\n\n")
iex> summary = llm_query.("Synthesize these analyses into a final report with root causes and recommendations:", combined)

Tips

• Start with Enum.take(lines, 10) to understand data structure before chunking
• Use shell("wc -l /path/to/file") to check file size first
• For code analysis, chunk by module/function rather than line count
• Keep chunk size at 300-800 lines for best sub-LLM accuracy
• Use write_file to save the final synthesis report