EDIP
Exploration’s reach at retrieval’s cost.
The org-scoped, multi-modal, agent-paced memory and decision pipeline underneath all Metantel harnesses. Connects code, docs, tickets, and media into a pre-linked Connectome served in judged rounds by our Librarian Retrieval Harness—and wires any data into plain-English workflows.
EnterpriseRAG-Bench (ahead of OpenAI, Amazon, Azure, Google)
Ranked first on conditional enterprise questions (95.6 score)
Net cost on codebase Q&A ($0.32 → $0.16 per question, 37 of 40 cheaper)
From git push to searchable across every agent—no disk checkout needed
The Pain Point, Our Solution, and Why Metantel.
Two ways to feed an agent today. Both break.
Exploratory agents (Claude Code, Codex) search and follow file names across hops—high reach, but massive token re-read cost, and company docs have no file names to follow. Precision retrieval and raw MCP search tools (Glean, standard RAG, memory APIs) do one search and return one fixed bundle—cheap, but blind past one hop.
Connect a Harness, not a raw MCP: Connectome + Librarian Retrieval Harness
A raw MCP tool just dumps unjudged data into your agent’s context window. EDIP connects a stateful Librarian Retrieval Harness: it builds multi-hop links ahead of time (the Connectome), judges every candidate step, and serves whole, verified documents in agent-paced rounds until the agent has enough.
Top-tier multi-hop accuracy at half the token cost, plus plain-English Workflows
EDIP ranks 4th of 26 overall and #1 on conditional questions on EnterpriseRAG-Bench over 511,962 records, cuts Q&A cost by 51% ($0.32 → $0.16) with identical fact recall (93.1% vs 91.9%), and includes EDIP Workflows to turn any data stream into an API endpoint.
Ask architectural "why" questions and get the exact memo, ticket, and commit thread in one round—without even checking out the repository on disk.
Automatic 4-layer memory (Live, Sessions, Knowledge, Team) means decisions made in Slack, docs, or a teammate’s agent session immediately inform your agent.
Org-ladder access control ensures facts only flow according to team clearance. Build visual plain-English decision boards with EDIP Workflows and publish them as REST/Python endpoints.
Why Connecting a Harness Beats Calling a One-Shot MCP
A retrieval pipeline is only as good as its bottleneck: finding evidence across multiple hops, and delivering it to the model whole. When you wire a raw MCP search tool to an agent, one-shot search results force the coding agent to pay tokens to sort through noise, while one-shot summarized answers cut off crucial evidence before the strong model ever sees it.
EDIP is a stateful, agent-paced Retrieval Harness. It serves judged whole documents in rounds. A single session remembers what it has already served, and the agent pulls rounds until it has enough—while always being told what evidence remains.
Multi-Layered Memory Distribution
| Memory Layer | Horizon | How It Serves Your Team |
|---|---|---|
| Live | Short-term | Every agent’s active intent and uncommitted edits, per repository and branch, held in live memory for immediate coherence. |
| Sessions | Experience | Every session’s context and key instructions stored in EDIP so important patterns persist and sessions graduate into AI Workers. |
| Knowledge | Long-term | Code, docs, tickets, and notes. The Connectome links them automatically ahead of time; the Librarian Retrieval Harness judges every step. |
| Team | Organisational | Facts written along the org ladder. Datafeeds are strictly scoped to team-level access (302/302 allowed facts delivered, 0/573 leaked in audit tests). |
EDIP Workflows (Beta): Any Data, Wired Into Decisions in Plain English
- Wire parts on a visual breadboard: connect data sources, plain-English decision nodes, shaping rules, results, and alerts—or describe the entire board in a single sentence.
- Every decision is an audited question: asked of every incoming item (notes, reports, code, or images), returning a calibrated confidence score and the exact evidence behind each answer.
- Publish as an endpoint in one click: any internal application or AI agent can call a published workflow via REST API or the Python SDK.
- Same memory underneath: the exact EDIP engine that feeds Claude Code also runs operational decisions across multi-modal enterprise data.
Internal & Public Benchmark Evidence
On EnterpriseRAG-Bench (500 complex queries across 511,962 company records, scored against the 28 Aug 2026 leaderboard), EDIP ranked 4th overall—ahead of OpenAI, Amazon, Azure, and Google search—and #1 on conditional questions. In internal codebase Q&A tests (40 questions, no edits), EDIP was cheaper on 37 of 40 runs, cutting average cost from $0.32 to $0.16 (−51% net) while maintaining 93.1% vs 91.9% fact accuracy—even when running with no git checkout on disk.
How EDIP connects to the Metantel stack
Every Metantel product shares one underlying memory and governance layer. Context, verification proofs, and access rules flow between tools automatically.
Delivers agent-paced evidence in rounds so coding sessions use 47% fewer tokens.
Explore Flow Console →Supplies product specs and runbooks so QA tests against what the business intended, not just the code.
Explore Artel QA →Provides the organizational access model and context that Agentify uses to govern web-app endpoints.
Explore Agentify →