RLM Research Engine

The RLM (Recursive Language Model) research engine lets you ask questions about materials science and get answers grounded in real knowledge graph data. Based on arxiv:2512.24601v2 by Zhang, Kraska, and Khattab.

How It Works

Unlike traditional RAG (retrieve then generate), RLM uses a REPL loop:

1. Your question goes to an LLM
2. The LLM writes Python code to explore the knowledge graph
3. The code executes on the runtime with injected tools
4. Real results come back as stdout
5. The LLM sees the data and writes more code or calls FINAL(answer)

Key difference: The LLM iterates. It writes one code block, observes the output, adapts its strategy, and repeats until it has enough data to answer.

API

POST /api/v1/knowledge/research/query

Request:

{
  "query": "Find materials containing nickel",
  "depth": 0
}

Depth controls cost:

• depth: 0 — free, uses local knowledge graph only
• depth: 1+ — costs money, enables web search (Semantic Scholar, arXiv, PubMed)

Response: Server-Sent Events (SSE) stream:

Event	Description
started	Session ID, mode, max_depth
reasoning	LLM’s thought process + code blocks
repl_exec	Code block being executed
repl_result	stdout/stderr from execution
answer	Final answer (from FINAL() call)
complete	Metrics: cost, graph_queries, llm_calls

Available Tools in REPL

# Search entities by name (substring match)
results = graph_search("nickel", limit=10)
# → [{"name": "Nickel Carbonyl", "entity_type": "CHM", ...}, ...]

# Get connected entities (1-hop traversal)
neighbors = get_neighbors("Ti-6Al-4V", limit=15)
# → {"nodes": [...], "edges": [...]}

# Semantic similarity search
docs = vector_search("creep resistance of superalloys", limit=5)
# → [{"content": "...", "similarity": 0.87}, ...]

# Web search (depth >= 1 only)
papers = web_search("NASA thermal protection materials", limit=5)
# → [{"title": "...", "abstract": "...", "doi": "..."}, ...]

# Query a sub-LLM
answer = llm_query("Summarize: {data}")
# → string response

# Mark final answer
FINAL("Your answer here")

PRISM CLI

prism research "Find materials containing nickel" --depth 0 --json

Example Output

A query for “nickel compounds” returns real data from the knowledge graph:

Found 20 entities containing 'nickel'
  Nickel Carbonyl [CHM]
  Nickel Hexacyanoferrate [CHM]
  Nickel Phosphide [CHM]
  Nickel Sulfide [CHM]
  ...

The LLM then explores deeper — searching for Ni-prefix entities, getting neighbors, running semantic searches — building up a comprehensive answer from actual graph data. Zero hallucination.