💻 computer science

Aethel: A Reproducible Graph-Retrieval Framework for Multi-Hop Financial Diligence

The paper introduces Aethel, a reproducible graph-retrieval framework that combines bipartite Personalized PageRank with coreference-aware teleportation and specialist agents to effectively synthesize fragmented financial disclosures for multi-hop diligence, demonstrating superior multi-hop recall and graceful degradation on large corpora compared to dense retrieval methods.

Original authors: Krish Sapru

Published 2026-07-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Krish Sapru

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a massive, multi-part mystery, but the clues are scattered across thousands of pages of boring, dense documents. This is the world of Information Retrieval, a branch of computer science dedicated to teaching machines how to find the right needle in a haystack. For years, computers have been great at finding needles that look exactly like the one you described (keyword matching) or needles that feel "similar" in meaning (using mathematical maps called embeddings). But they often struggle when the answer requires connecting two dots that are far apart, like realizing that a company's name on page 10 is the same entity mentioned as "the fund" on page 500. This is called multi-hop reasoning, and it's crucial for high-stakes tasks like financial investing, where missing a single link in the chain can cost millions. The big question researchers are asking is: Can we build a smarter system that doesn't just look for keywords or similar words, but actually understands the connections between pieces of information, like a detective following a trail of breadcrumbs?

Enter Aethel, a new framework designed to tackle this exact problem in the high-pressure world of private equity finance. Think of Aethel not as a simple search bar, but as a team of specialized detectives working together. Instead of just reading documents one by one, Aethel first builds a giant, invisible web (a graph) where every important name (like a company or a person) is a node, and every paragraph of text is another node. These nodes are connected by lines whenever a name appears in a paragraph. When a user asks a complex question, Aethel uses a mathematical trick called Personalized PageRank—imagine dropping a drop of ink at the starting clue and watching it flow through the web, coloring the most relevant paragraphs based on how easily the ink can reach them. This allows the system to "hop" from one document to another, bridging gaps that other systems miss.

However, the paper's most surprising finding is that this fancy web-based detective work isn't always the winner. The researchers tested Aethel in two very different scenarios. First, in a "closed pool" where the system only had to choose from a tiny, pre-selected list of 10 to 20 paragraphs, Aethel's web-based approach was excellent. It successfully found all the necessary clues 100% of the time on one test set and 88.5% on another, even when the clues were scattered. This is great because in a real investigation, you want all the evidence, not just the single best-looking piece.

But the story changes when the system faces the real world. The researchers then tested Aethel on a massive, open library of 4,123 real financial documents, including SEC filings and earnings reports. Here, the "fancy" graph system didn't beat the old-school, simple method of matching exact words (known as BM25). In fact, the simple word-matching method was actually better at finding the right answers. While Aethel's graph method was better than the "similar meaning" systems at finding multi-step connections, it still couldn't outperform the simple keyword matcher. The authors suggest that on keyword-heavy financial texts, the simple method is just too strong to beat, and the graph's advantage gets diluted when the library of documents gets too big. They found that combining the two methods (the graph and the keyword matcher) gave a tiny, barely noticeable improvement, but it wasn't statistically significant enough to call a clear victory.

Ultimately, the paper suggests that while graph-based systems like Aethel are powerful tools for connecting dots in small, controlled groups of documents, they don't automatically replace the reliable, old-fashioned keyword search when dealing with huge, messy libraries of financial text. The most dramatic discovery was that the "similar meaning" systems actually fell apart as the document library grew, while the simple keyword method stayed steady. So, for now, the best strategy might be to use the simple method as the backbone and perhaps use the graph system as a helpful sidekick, rather than expecting the graph to do all the heavy lifting alone.

Technical Summary: Aethel – A Reproducible Graph-Retrieval Framework for Multi-Hop Financial Diligence

Problem Statement

Secondary private equity (PE) transactions require analysts to synthesize fragmented, unstructured financial disclosures under compressed timelines. Critical metrics and their entity anchors are often distributed across disjoint documents (e.g., LP letters, capitalization structures, SEC filings) with negligible lexical overlap. Standard Dense RAG systems, which rely on cosine similarity in embedding spaces, falter in these scenarios because they cannot bridge topological gaps where a metric and its anchoring entity are separated by dozens of pages or distinct document types. While recent Graph-Augmented RAG (GraphRAG) systems offer promise, existing frameworks often remain brittle in document-intensive corporate diligence domains where entity names appear in multiple surface forms and critical data is topologically isolated.

Methodology: The Aethel Framework

Aethel is a graph-retrieval framework designed to resolve relational multi-hop queries by modeling corpora as bipartite entity-passage graphs. It pairs a retrieval engine with an orchestrated agent swarm.

1. Bipartite Graph Representation

The system ingests documents and constructs an undirected bipartite graph $G = (V_P \cup V_E, E)$ , where:

$V_P$ represents discrete text passages (document chunks).
$V_E$ represents unique extracted entity nodes.
Edges $E$ map undirected co-occurrence between entities and passages.

2. Personalized PageRank (PPR) Retrieval

Instead of relying solely on direct dense similarity, Aethel uses PPR to propagate relevance through the graph.

Seeding: Query entities ( $E_q$ ) serve as seed nodes in a personalization vector $s$ .
Diffusion: A random walk is performed using the transition probability matrix $W$ derived from the adjacency matrix. The steady-state probability vector $p$ is solved via power iteration: $p^{(t+1)} = \alpha s + (1 - \alpha)Wp^{(t)}$ .
Mechanism: The teleportation term ( $\alpha$ ) keeps the walk anchored to query entities, while the transition term allows evidence to diffuse across neighboring passages and related entities, bridging gaps where lexical overlap is low.

3. Bipartite Coreference Teleportation (BCT)

To address the issue of entity names appearing in multiple surface forms (e.g., "NorthRiver" vs. "NorthRiver IV LP"), Aethel introduces a BCT layer. This layer expands the PPR personalization vector via alias matching and substring overlap. This increases the probability mass assigned to relevant passages that share the same underlying entity but differ in surface form, prioritizing hit-rate coverage (finding all necessary supporting passages) over top-1 precision.

4. Specialist Swarm Architecture

Retrieved evidence is routed to a downstream swarm of specialist agents (e.g., Liquidity Specialist, Valuation Auditor, Diligence Auditor). The graph-walk step acts as a topological filter, constraining the context window for each specialist to the highest-relevance passages while preserving cross-document audit trails. The system synthesizes findings into an investment committee memo.

Key Contributions

Bipartite Coreference Teleportation (BCT): A formalized correction layer that improves Hit Rate@5 (HR@5) by expanding entity mentions in the PPR vector. The authors note this trades precision@1 for coverage, a favorable exchange for multi-passage diligence where all supporting evidence must be surfaced.
Graph-Filtered Specialist Swarm: An architecture description coupling bipartite PPR retrieval with a parallel agent swarm. The graph serves as a structural constraint to prevent contextual fragmentation, though agent-level performance is not separately benchmarked in this work.
Open-Corpus Financial Evaluation: The construction of a 4,123-chunk corpus of real financial disclosures (SEC filings, earnings materials, secondary-market reports) with 40 manually annotated queries. This evaluation reveals a critical scale-dependence finding regarding graph-RAG efficacy.

Empirical Results

Closed-Pool Evaluation (MuSiQue & 2WikiMultiHopQA)

Evaluated on 200-question random samples with a closed candidate pool (10–20 paragraphs per question):

Performance: The BCT-enhanced graph retriever achieved HR@5 of 88.5% on MuSiQue and 100.0% on 2WikiMultiHopQA.
Trade-off: This improvement in coverage came at the cost of precision@1 (HR@1 dropped from 63.0% to 57.0% on MuSiQue).
Context: In this small-pool setting, the dense bi-encoder baseline also performed well (HR@5 ~96%), but the graph method ensured all supporting passages were captured, which is critical for downstream synthesis.

Open-Corpus Evaluation (Real Financial Disclosures)

Evaluated on a 4,123-chunk corpus with 40 queries (20 single-hop, 20 multi-hop):

Lexical Dominance: A strong lexical baseline (BM25) outperformed all learned and graph-based methods on HR@5 (0.700 vs. 0.600 for Aethel-NER3 and 0.450 for the dense bi-encoder).
Graph vs. Dense: The graph retriever (Aethel-NER3) significantly outperformed the dense bi-encoder on multi-hop queries (HR@5 0.600 vs. 0.450), demonstrating that graph topology recovers associative recall that dense embeddings miss at scale.
Fusion Results: Reciprocal Rank Fusion (RRF) of BM25 and Aethel-NER3 yielded an MRR of 0.479 compared to BM25's 0.468. However, a paired bootstrap test (N=20, 95% CI [−0.163, +0.192]) showed this difference does not clear zero, indicating the gain is suggestive rather than statistically significant.
Entity Index Quality: The results were highly sensitive to entity indexing. A naive regex-based graph (Aethel-Reg) performed poorly (HR@5 0.400). A corrected version using substring adjacency and filtered entities (Aethel-NER3) was required to recover the reported multi-hop performance.

Scale Sensitivity

A scaling experiment (100 to 4,123 chunks) revealed:

Dense Retrieval Collapse: Dense retrieval performance collapsed as the corpus grew (HR@5 dropped from 0.840 at 100 chunks to 0.450 at full scale).
BM25 Robustness: BM25 remained robust, degrading only slightly (0.800 to 0.700).
Graph Advantage: The graph retriever overtook dense retrieval starting at 500 chunks, with the gap widening monotonically. However, the graph retriever never overtook BM25 at any tested scale.

Significance and Claims

The paper's central empirical finding is that graph-RAG's multi-hop advantage is pool-size dependent.

In closed candidate pools (small scale), the graph approach, particularly with BCT, provides superior coverage of supporting evidence compared to dense retrieval.
In open-corpus settings (large scale, keyword-dense financial text), the graph retriever's advantage over dense retrieval is clear, but it does not extend to beating a strong lexical baseline (BM25).
The authors conclude that while graph topology helps recover multi-hop associations that dense embeddings miss, the "teleport-and-diffuse" dynamics of PPR spread probability mass too thinly in large, heterogeneous graphs, allowing keyword-anchored lexical baselines to dominate recall.
The paper positions the gap between closed-pool and open-corpus results as the primary takeaway: the multi-hop advantage of graph-RAG is substantially attenuated at open-corpus scale on financial text. The authors report the fusion results honestly as preliminary and not statistically significant at the current sample size, calling for larger, multi-annotator benchmarks to draw conclusive results.