Ask a specific question
Ask about an ESG topic, regulatory requirement, company disclosure or comparison. For example: “Compare Infosys and Reliance on climate-related performance.”
The system embeds your query with Supabase gte-small, fuses semantic vector similarity with PostgreSQL full-text search, and sends only the retrieved evidence to the generation layer.
Use ESG RAG Intelligence to interrogate the indexed BRSR framework, BRSR Core guidance and corporate ESG disclosures. The tool is designed for evidence-led research, not as a replacement for the underlying reports.
Ask about an ESG topic, regulatory requirement, company disclosure or comparison. For example: “Compare Infosys and Reliance on climate-related performance.”
Use the source and ESG-theme filters when you want to narrow the search. For comparisons, name both companies and ask specifically about reporting year, methodology, scope or boundary differences.
Material claims in the answer cite evidence IDs such as [E1] and [E2]. Review the Retrieval Trace, check the PDF page and use “source ↗” to inspect the original disclosure.
No synthetic ESG metrics are seeded into the source database. The ingestion function fetches only the allow-listed official document URLs stored in Supabase.
The database contains ten expected-source evaluation questions. Run them against the live pgvector index after initialization.
| Question | Expected source(s) | Top retrieved source(s) | Hit@5 |
|---|
Document chunks are rows in PostgreSQL and embeddings are stored in a native vector(384) column. HNSW accelerates semantic nearest-neighbour search.
A generated tsvector column and GIN index preserve exact ESG terminology such as Scope 3, BRSR Core and assurance.
Semantic and full-text result lists are merged using RRF so neither exact terminology nor semantic intent dominates by default.
The browser receives only the public Supabase key. Database writes and retrieval logging happen inside JWT-protected Edge Functions using the service role.
The generation endpoint is instructed to use only [E#] evidence, identify boundary/year differences and abstain when retrieval is insufficient.
Queries, filters, returned chunk IDs and retrieval latency are recorded server-side for future evaluation and debugging.