- Ingest data
- Explanation: The current GraphRAG query routing system is live and fully working. However, in the current sparse state of the database, nodes might be genuinely unrelated but superficially connected through random articles. Real, high-quality logical and causal connections between entities can only emerge once a massive volume of articles is fetched, extracted, and ingested into the graph database, providing the necessary density for graph traversals to become highly accurate.
- Refine LLM Entity Extraction
- Explanation: The current extraction logic often pulls overly generic terms (e.g., "Government", "PM", "Budget") or temporal words (e.g., "July", "September"). This leads to false equivalencies during semantic merging (e.g., improperly linking unrelated events just because they happened in the same month, or conflating "UK Government" with "Indian PM" under generic terms). The extraction prompt needs strict constraints to filter out generic/temporal terms and only extract highly specific, disambiguated proper nouns.
- Autonomous / Agentic Ingestion Feedback Loop
- Explanation: Currently, ingestion requires a lot of human input (Initial, Targeted, and Requested methods). To make the system truly agentic, a feedback loop needs to be established with the database. The system should autonomously analyze the graph database to identify sparse areas or knowledge gaps, determine what sort of topics/entities need more context, and automatically execute recommended ingestion pipelines. This will reduce manual intervention and allow the graph to organically densify itself.
Folders and files
| Name | Name | Last commit date | ||
|---|---|---|---|---|