Interlock extracts board-interlocks, audit relationships, related-party transaction trails and promoter-pledge chains from NSE/BSE annual reports — loads them into a TigerGraph knowledge graph — then runs RAG, GraphRAG and Agentic GraphRAG side by side on the same questions so you can see exactly where the graph changes the answer.
The problem
A single annual report rarely tells the full story. Risk emerges from relationships that span multiple documents, companies and fiscal years — exactly the shape that plain vector search cannot follow.
Board interlocks
A director sitting on five boards simultaneously — including one company that later faced a SEBI order — is invisible inside any single filing. You need to traverse the graph.
Related-party trails
Material transactions routed through subsidiary chains obscure beneficial ownership. Answering 'who ultimately received the payment' requires multi-hop traversal, not keyword matching.
Auditor independence
An audit firm auditing the same promoter-group entities for 15 years is a red flag that only appears when you join auditor records across companies and fiscal years.
Promoter pledge cascades
A promoter pledging shares in one entity to fund acquisition in another creates a risk chain that plain RAG cannot reconstruct — it has no notion of traversal depth.
Temporal changes
A director who resigned the quarter before a regulatory action is not suspicious in any one document. It becomes significant only when events are joined across time.
Why RAG fails here
RAG returns the top-K most similar chunks and asks the LLM to reason. It cannot aggregate, cannot join two facts from different PDFs, and cannot follow a chain of relationships.
The graph
Every extracted fact is an edge carrying a verbatim source quote and the PDF page it came from. The same graph powers all three pipelines.
InterlockV2 · live TigerGraph graph · 5 companies · FY2021–24
Tata Steel
Company
DIRECTOR_OF ×28
SUBSIDIARY_OF ×1
N. Chandrasekaran
Person
DIRECTOR_OF ×28
AUDITED_BY ×8
Tata Motors
Company
S R B C & CO LLP
AuditFirm
Bajaj Finance
Company
Bajaj Finserv
Company
18,994
Chunks with embeddings
3,240
MENTIONS edges
26
Source documents
The solution
All three share the same evidence budget, the same answer prompt and the same verifier. The only difference is how they retrieve evidence.
Embeds the question with all-MiniLM-L6-v2, searches the vector index for the top-40 chunks, fuses with BM25 keyword scores via reciprocal rank fusion, fits a 4,000-token evidence budget, and makes one LLM call.
A helper LLM call extracts entity mentions and relation types from the question. Those entities are linked to graph nodes, expanded up to 2 hops via an installed GSQL query, and the resulting triples + linked chunks are ranked and passed to the answer call.
A tool-calling loop runs under hard step, token and wall-clock budgets. The model picks from search_text, find_entity, neighbors, expand_hop, graph_query and calculate. After the loop, a separate verifier checks each claim against the evidence.
By the numbers
5
Companies
TATAMOTORS · TATASTEEL · BAJFINANCE · BAJAJFINSV + 1
26
Annual reports
FY2021-22 through FY2023-24
18,994
Text chunks
Embedded with sentence-transformers
60%
GraphRAG accuracy
dev split · 5 questions · live graph
100%
RAG accuracy
test split · 12 questions
40%
Agent accuracy
dev split · 5 questions · no-graph baseline
28
Board director edges
DIRECTOR_OF with provenance
All figures from stored evaluation runs. 17 questions total, 3 of 6 planned categories. The 100% RAG interval is degenerate at n=12 — read it as a working baseline, not a guarantee.
Architecture
A modular pipeline that turns public PDFs into a knowledge graph, then measures exactly where each retrieval strategy wins and loses.
Explore the results
The dashboard shows per-category accuracy with bootstrap CIs, cost and latency trade-offs, a full failure breakdown, and a side-by-side inspector for every question.