Synthesis & Super-Synthesis

Synthesis is where the knowledge graph comes alive — an AI agent navigates the graph, follows evidence trails, and weaves everything into a coherent, attribution-grounded research document.

Synthesis

The Synthesizer Agent

The SynthesizerAgent is a LangGraph-based agent that explores the knowledge graph to answer a research question. It operates with an exploration budget — a configurable limit on how many nodes it can visit, controlling investigation depth without restricting search.

Navigation tools

The agent has access to tools for exploring the graph:

Tool	Purpose	Budget Cost
get_node	Load a node's definition, edge count, fact count	1 unit per unvisited node
get_node_facts	Retrieve all facts for a node with source attribution	Free if already visited
get_node_dimensions	Load multi-model analyses for comparison	Free if already visited
search_graph	Find nodes by text search	Free
finish_synthesis	Submit the final document	Free

The agent decides which nodes to visit, which edges to follow, and when it has gathered enough evidence to write.

Synthesis principles

The agent follows strict principles when writing:

1. Answer first, evidence second — Lead with insight; facts are building blocks, not the conclusion
2. Attribution-grounded — Every claim connects to specific facts and sources. No bare assertions.
3. Radical source neutrality — No source gets automatic credibility based on institutional prestige. All receive equal scrutiny.
4. Preserve perspectives — When evidence supports multiple viewpoints, each gets dedicated coverage
5. Stakeholder analysis — For every claim, consider who benefits from belief in that claim
6. Citation discipline — Inline fact citations rendered as clickable links to original sources

Convergence-aware writing

The agent adapts its writing based on convergence scores:

Convergence	Treatment
> 0.7	Present as established, well-supported conclusion
0.4 - 0.7	Note uncertainty, present evidence from both sides
< 0.4	Present competing perspectives explicitly, let evidence speak

Output

A synthesis produces:

• Markdown document with inline fact citations
• Confidence score for the overall synthesis
• Cited facts and nodes — which evidence was used
• Divergences — where the synthesis encountered unresolved disagreements
• Subgraph — the nodes and edges traversed, for UI visualization

Super-Synthesis

When synthesis isn't enough

Some research questions are too broad or multi-faceted for a single synthesis agent. Super-synthesis solves this by orchestrating multiple synthesis agents working in parallel on different aspects of the question.

The Super-Synthesizer Agent

The SuperSynthesizerAgent is a meta-level agent that:

1. Reads sub-synthesis documents — Reviews the output of multiple synthesis runs
2. Identifies connections — Finds themes, patterns, and relationships across syntheses
3. Searches for gaps — Queries the graph for additional context that bridges between sub-topics
4. Produces a meta-narrative — Combines everything into a unified, cross-domain document

Super-synthesis tools

Tool	Purpose
read_synthesis	Load the full text of a sub-synthesis document
get_synthesis_nodes	List all nodes referenced in a synthesis
search_graph	Find additional connecting nodes
get_node	Load node details for cross-referencing
finish_super_synthesis	Submit the combined document

The workflow

1. Reconnaissance — The super-synthesizer plans scopes covering different angles of the topic
2. Parallel dispatch — Multiple synthesizer agents are launched, each investigating one scope
3. Collection — Sub-synthesis documents are gathered as they complete
4. Integration — The super-synthesizer reads all sub-syntheses, resolves overlaps, finds connections, and fills gaps
5. Meta-narrative — A unified document is produced that weaves all perspectives together

Example

For a broad question like "What causes cancer?", the super-synthesizer might plan:

• Agent 1: Genetic factors and hereditary predisposition
• Agent 2: Environmental and occupational exposure
• Agent 3: Lifestyle factors (diet, exercise, substance use)
• Agent 4: Emerging research (epigenetics, microbiome, immune system)
• Agent 5: Treatment implications and prevention strategies

Each agent produces a focused synthesis. The super-synthesizer then combines these into a comprehensive document that reveals cross-domain connections — like how genetic predisposition interacts with environmental exposure, or how lifestyle factors modulate immune responses.

The result is a research document that no single synthesis run could produce, grounded in the same evidence-first, attribution-based approach that underlies all of Knowledge Tree.