A Knowledge Graph-Grounded Multi-Agent Framework for Cross-Methodology Innovation Analysis and Patent Claim Generation
Companion code for the paper:
IdeaForge: A Knowledge Graph-Grounded Multi-Agent Framework for Cross-Methodology Innovation Analysis and Patent Claim Generation https://arxiv.org/abs/2605.13311
Most AI innovation tools apply a single methodology (TRIZ, Design Thinking, or SCAMPER) in isolation and discard intermediate reasoning. IdeaForge is different:
- Runs multiple specialist agents — TRIZ, Design Thinking, SCAMPER — each writing structured nodes and edges to a persistent FalkorDB knowledge graph
- A prior art agent searches arXiv for related work and populates PriorArt nodes
- An embedding synthesis agent uses sentence-transformer cosine similarity to detect claims independently derived by multiple methodologies — these are the strongest patent candidates
- InnovationScore ranks all claims using a weighted formula combining convergence, diversity, strength, and prior art challenge count
- A patent agent drafts structured claims grounded in the KG subgraph
- A visualizer generates interactive HTML and static PNG graph images for demos and papers
- An MCP server exposes KG tools to external agents
The central novel contribution: the CONVERGENT edge — connecting claims independently supported by multiple methodologies. Claims with high convergent count and methodology diversity are the strongest patent candidates.
Raw idea (text)
|
v
Problem node -> FalkorDB KG
|
+---> TRIZAgent -> Contradiction, Principle, Claim nodes
|
+---> DesignThinkingAgent -> UserNeed, Claim nodes
|
+---> SCAMPERAgent -> Transformation, Claim nodes
|
+---> PriorArtAgent -> PriorArt nodes (arXiv search)
|
v
EmbeddingSynthesisAgent -> CONVERGENT edges (cosine similarity)
|
v
InnovationScore -> ranked claim list
|
v
PatentAgent -> Patent draft grounded in KG subgraph
|
v
MCP Server -> Exposes KG tools to external agents
Visualizer -> HTML + PNG graph image
Nodes:
Problem (statement, domain)
Contradiction (improving, worsening)
Principle (name, triz_number, description)
UserNeed (persona, job_to_be_done, pain_level)
Transformation (scamper_type, description)
Analogy (source_domain, mechanism) [reserved for future biomimicry agent]
PriorArt (title, source, similarity)
Claim (text, methodology, strength)
Edges:
(Problem)-[:HAS_CONTRADICTION]->(Contradiction)
(Contradiction)-[:RESOLVED_BY]->(Principle)
(Principle)-[:SUPPORTS]->(Claim)
(UserNeed)-[:MOTIVATES]->(Problem)
(Transformation)-[:GENERATES]->(Claim)
(Analogy)-[:INSPIRES]->(Claim)
(PriorArt)-[:CHALLENGES]->(Claim)
(Claim)-[:CONVERGENT {count}]->(Claim) <-- key novel contribution
Claims are ranked by a weighted formula:
InnovationScore(c) = 0.4 * convergent_count
+ 0.3 * methodology_diversity
+ 0.2 * claim_strength
- 0.1 * prior_art_challenge_count
Where:
convergent_count— CONVERGENT edges on this claim (cross-methodology support)methodology_diversity— distinct methodologies independently supporting the claimclaim_strength— fixed by methodology: TRIZ=0.7, DesignThinking=0.65, SCAMPER=0.6prior_art_challenge_count— PriorArt nodes challenging this claim
The claim with highest InnovationScore becomes the primary independent claim in the patent draft.
ideaforge/
├── kg/
│ ├── schema.py # Node/edge types and Cypher templates
│ └── graph.py # FalkorDB graph operations
├── agents/
│ ├── methodology_agents.py # TRIZ, DesignThinking, SCAMPER agents
│ ├── prior_art_agent.py # arXiv search -> PriorArt nodes
│ ├── embedding_synthesis.py # Cosine similarity convergence detection
│ ├── innovation_score.py # InnovationScore computation and ranking
│ └── patent_agent.py # Patent claim drafting from KG
├── mcp_server/
│ └── server.py # MCP server exposing KG tools
├── visualize.py # pyvis HTML + networkx PNG graph visualization
├── run_experiments.py # Multi-domain evaluation + threshold sensitivity
├── ideaforge.py # Main pipeline entry point (8-step)
├── docker-compose.yml
└── requirements.txt
# Start FalkorDB
docker-compose up -d
# Install dependencies
pip install -r requirements.txt
# Install Ollama + TinyLlama
# https://ollama.com
ollama pull tinyllama
ollama servepython ideaforge.py --idea "A voice-first legal assistant in Hindi for rural India"python ideaforge.py \
--idea "A voice-first legal assistant in Hindi for rural India" \
--domain "legal technology" \
--output patent_draft.txt
python visualize.py --staticpython ideaforge.py --idea "your idea" --dry-runOLLAMA_MODEL=llama3.2 python ideaforge.py --idea "your idea"
OLLAMA_MODEL=mistral python ideaforge.py --idea "your idea"FALKORDB_PORT=6380 python run_experiments.pyStep 1: Ingest idea into knowledge graph
Step 2: TRIZ contradiction analysis
Step 3: Design Thinking user needs analysis
Step 4: SCAMPER transformations
Step 5: Prior art search (arXiv)
Step 6: Embedding-based convergence detection
Step 7: InnovationScore ranking
Step 8: Patent draft from KG subgraph
InnovationScore Report
============================================================
Rank Score Conv Div PA Claim
------------------------------------------------------------
1 0.500 2 2 0 [TRIZ] A method for resolving the contradiction
between accessibility and complexity of legal
language, applying Segmentation and Preliminary
Action principles to voice-based delivery
2 0.310 1 1 0 [DesignThinking] A user-centred system enabling
rural citizens to query legal rights in Hindi
via voice, without needing a lawyer
3 0.220 0 1 0 [SCAMPER] A transformed approach substituting
text-based legal interfaces with voice-first
interaction, adapting medical triage dialogue
patterns to legal question routing
Convergent pairs detected: 3
TRIZ + DesignThinking: 0.837
TRIZ + SCAMPER: 0.817
DesignThinking + SCAMPER: 0.819
Knowledge Graph: 16 nodes, 10 edges
| Use Case | Domain | Nodes | Conv. pairs | Top score |
|---|---|---|---|---|
| Voice-first legal assistant (Hindi) | Legal tech | 16 | 3 | 0.500 |
| Sepsis early warning (wearables) | Healthcare | 16 | 1 | 0.353 |
| Adaptive tutoring for dyscalculia | EdTech | 16 | 3 | 0.500 |
| Drone crop disease detection | Agriculture | 16 | 3 | 0.467 |
| Sign language interpretation | Accessibility | 16 | 3 | 0.500 |
# Interactive HTML (for demos)
python visualize.py
# Static PNG (for paper figure)
python visualize.py --staticCONVERGENT edges are shown in pink — they are the visual centrepiece of the graph and the paper's core contribution.
| Tool | Description |
|---|---|
get_all_claims |
All claims sorted by strength |
get_convergent_claims |
Claims with cross-methodology support |
get_strongest_claims |
Top 5 patent candidates |
get_kg_summary |
Node count summary |
add_claim |
Add a claim to the KG |
python mcp_server/server.py| System | Methodology | Persistent KG | Cross-methodology synthesis | InnovationScore | Patent draft |
|---|---|---|---|---|---|
| AutoTRIZ (2024) | TRIZ only | No | No | No | No |
| TRIZ Agents (ICAART 2025) | TRIZ only | No | No | No | No |
| LLM+TRIZ Patent (2026) | TRIZ only | No | No | No | Yes |
| IdeaForge | TRIZ + DT + SCAMPER | Yes | Yes (embeddings) | Yes | Yes (KG-grounded) |
The central argument: innovation methodologies can be interpreted as heterogeneous reasoning operators acting over a shared persistent innovation graph. Cross-methodology convergence — the same claim emerging independently from multiple operators — is a principled signal of non-obviousness.
- LLM quality affects agent outputs — TinyLlama produces basic results; larger models produce richer claims
- Convergence detection uses semantic similarity, not true logical equivalence
- Prior art search is limited to arXiv — patent database integration is future work
- No legal validation — this is a research prototype, not a patent filing tool
- InnovationScore weights are heuristic — formal novelty cannot be guaranteed
@article{bose2026ideaforge,
title={IdeaForge: A Knowledge Graph-Grounded Multi-Agent Framework for Cross-Methodology Innovation Analysis and Patent Claim Generation},
author={Bose, Joy},
journal={arXiv preprint arXiv:2605.13311},
year={2026}
}