Better retrieval. Connected context.
Measured in better answers.
Quality score out of 100
Same model, agent, and prompts
Points in answer quality
All question banks · 0 to 100 model judgments with source review.
17 paired questions
Quality: Metal / file platform
Time to answer
Tokens used
15 paired questions
Quality: Metal / file platform
Time to answer
Tokens used
19 paired questions
Quality: Metal / file platform
Time to answer
Tokens used
Higher quality in every bank. The widest quality gap was hard reconciliation: +13.0 points.
TESTED CONFIGURATION
Evaluation notes
Same model, agent, and questions through both MCP backends. Workflow quality averages factual accuracy and task completion. Overall mean latency: 228s with Metal vs 454s with the file platform. Tokens: 1.13M vs 1.90M. Mean tool errors: 2.7 vs 9.1. Model configuration and repeated-trial variance are not specified.
Metal combines retrieval with a Context Graph that connects documents, companies, people, deals, and activities. This evaluation compares that combined context backend with an enterprise file platform.
Connect a document to its company, deal, people, and activity history. Give the agent structured context alongside the evidence it retrieves.
The reported gains measure the combined backend. They do not isolate the contribution of retrieval or the graph individually.
The comparison holds the model, agent, and prompts constant. The context source changes: Metal or a leading enterprise file platform, each reached through MCP.
Roughly 100,000 real files, including CIMs, IC and underwriting materials, financing and valuation models, lender and portfolio reporting, plus pipeline and screening records. The same model and agent access each backend through MCP.
17 hard reconciliation tasks, 15 CIM discovery tasks, and 19 structured workflows, built from live deal work. Nineteen tasks cannot be answered from documents alone: they require joining documents to the firm’s pipeline records.
Every answer is graded from 0 to 100 against the source evidence, with each citation checked. The model, agent, and prompts are identical on both sides. Time and tokens are means per answer.
Explore the capabilities each bank evaluates, from reconciling financial evidence to completing connected deal workflows.
Tests whether an answer preserves financial definitions, distinguishes historical results from forecasts, reconciles source conflicts, and supports conclusions with evidence.
Tests whether an answer finds relevant investment materials, separates businesses from document copies, compares compatible metrics, and preserves uncertainty when evidence is incomplete.
Tests whether an answer connects documents to structured records, preserves scope and current state, and completes a usable workflow without overstating what the evidence establishes.
The same 51 questions, organized by lifecycle stage.
| Lifecycle stage | Questions | What the questions ask for |
|---|---|---|
| Sourcing, intake and screening | 29 | Sector screens, CIM receipt versus publication dates, codename collisions, screening readiness, sector routing, acquisition shortlists |
| Pipeline and add-on management | 4 | Closed and passed inventories, platform relationships, add-on overlap, opportunity coverage |
| Underwriting and IC | 4 | Add-on underwriting tables, capital intensity, IC postmortems, IC-readiness handoffs |
| Financing and capital structure | 3 | Financing chronologies, sources and uses, leverage and covenant reconciliation |
| Portfolio monitoring and reporting | 9 | EBITDA bridges, conflicting EBITDA versions, lender reporting, retention, valuation, re-underwriting |
| Exit and divestiture | 2 | Divestiture tracing, exit return versions |
Explore how Metal retrieves evidence and connects your firm’s context across systems.
Explore Metal with your team