Research Mode (RLM)

The RLM Research Engine is an AI agent that recursively decomposes complex research questions, explores the knowledge graph, searches academic databases, ingests papers, and synthesizes cited answers. Based on the Recursive Language Models paper (arxiv:2512.24601v2).

Unlike flat retrieval (RAG), the RLM agent decides what to query, inspects results, and can spawn independent sub-sessions for different aspects of a complex question. Each sub-session has its own tools and context. Results are synthesized bottom-up.

Endpoint

POST /api/v1/knowledge/research/query

Returns a Server-Sent Events (SSE) stream of steps.

Usage

curl -N -X POST -H "X-API-Key: $KEY" \
  -H "Content-Type: application/json" \
  "https://api.marc27.com/api/v1/knowledge/research/query" \
  -d '{"question": "What materials resist erosion in rocket nozzles?", "depth": 2}'

Depth Pricing

Depth	Cost	What it does
0	Free	Local graph + vector search only
1	$0.01	+ Web search (Semantic Scholar, arXiv, PubMed, OpenAlex) + DOI lookup + URL crawling
2	$0.05	+ Recursive decomposition into sub-sessions
3	$0.10	+ Deep recursion (sub-sessions can spawn sub-sessions)

Tools Available to the Agent

Tool	Depth	What it does
graph_search	0+	Fulltext search on entity names
get_neighbors	0+	1-hop graph traversal
vector_search	0+	Semantic similarity search
property_search	0+	Find materials by property (e.g., yield strength)
compute_run	0+	Submit DFT/ML job (MACE, QE, LAMMPS)
web_search	1+	Semantic Scholar + arXiv + PubMed + OpenAlex
doi_lookup	1+	CrossRef DOI resolution (metadata, references, funding)
crawl_url	1+	Crawl any URL and extract clean text
ingest_paper	1+	Extract entities from a paper into the graph permanently
sub_query	2+	Recursively decompose into independent sub-sessions
synthesize	all	Generate final answer with citations

Recursive Decomposition (depth 2+)

At depth 2 and above, the agent can break complex questions into independent sub-questions. Each sub-question runs as a separate research session with its own agent loop.

Question: "Compare Ti-6Al-4V and Inconel 718 for turbine blade applications"
  -> sub_query: ["Tensile and creep properties of Ti-6Al-4V at high temperature",
                  "Oxidation resistance of Inconel 718 above 800C"]
  -> Sub-session 1: searches graph, finds papers, ingests data on Ti-6Al-4V
  -> Sub-session 2: searches graph, web, ingests papers on Inconel 718
  -> Parent synthesizes both sub-answers into comparative analysis

Flywheel

Every web search result that gets ingested becomes permanent knowledge. The first query about a topic costs money (web search + LLM extraction). Every subsequent query finds it locally for free. The more the system is used, the smarter and cheaper it gets.

SSE Event Format

data: {"step": "reasoning", "data": {"turn": 0, "text": "I'll start with vector search..."}}
data: {"step": "tool_call", "data": {"tool": "vector_search", "query": "nozzle erosion"}}
data: {"step": "tool_result", "data": {"count": 5, "tool": "vector_search"}}
data: {"step": "decompose", "data": {"sub_queries": ["...", "..."], "reasoning": "..."}}
data: {"step": "sub_session_start", "data": {"sub_query": "...", "depth": 2}}
data: {"step": "sub_session_complete", "data": {"sub_query": "...", "answer": "..."}}
data: {"step": "answer", "data": {"text": "Based on the analysis..."}}
data: {"step": "complete", "data": {"metrics": {"graph_queries": 12, "cost_usd": 0.05}}}

Via PRISM CLI

prism query --platform --semantic "What materials resist nozzle erosion?"