Describe Plan -> Search Records -> Rank Matches -> Summarize -> Hand to Chemist

R&D Experiment Knowledge Miner

The agent searches scattered records, electronic notebook exports, scanned batch sheets, and project reports, to find experiments and routes similar to what a chemist is planning. It returns a ranked list with the conditions used and the outcomes, and a short summary. It is advisory only, the chemist decides what to run.

The Problem

The reaction was run three years ago. The chemist ran it again to find out.

A chemist plans a route for a new project and spends two days at the bench working out conditions. The same chemistry was run three years ago by someone who has since left, with the conditions written in a notebook that was scanned and filed. Nobody can search a scanned notebook, so the knowledge is effectively lost and the work is repeated. The old result would have saved the two days.

What Changes

Search the notebooks nobody can search, before the bench work starts.

The agent reads the electronic notebook exports, the scanned batch sheets with text recognition, and the project reports, and finds the experiments closest to what the chemist is planning. It returns a ranked list with the conditions and the outcomes and a short summary of what worked and what did not. The chemist uses it as a starting point, not a decision.

Humans stay in control: the agent retrieves and summarizes prior work, and the chemist decides what conditions to use and what to run.

HOW THE AGENT WORKS

TRIGGERLLMAPIDECISIONACTIONWAIT

TRIGGERChemist queries a planFires when a chemist describes a planned experiment or route.

APISearch experiment corpusSearches notebook exports, scanned sheets, and reports for similar work.

LLMRank and extract conditionsRanks the matches and pulls the conditions and outcomes from each.

DECISIONAny strong match?

LLMSummarize what workedWrites a short summary of conditions that worked and pitfalls seen.

ACTIONReturn results to chemistDelivers the ranked matches and summary to the chemist.

The agent does not design the experiment or decide the route. Your chemist does. The agent removes the manual hunt through old notebooks and repeated work.

HOW TO BUILD IT

Three levels of depth. Pick the one that matches where you are.

For the ops person who needs to understand what's involved and brief their team.

Experiment RecordsNotebooks, scanned sheets, reports

The AgentWorkflow engine + LLM

ChemistReads the matches and decides

WHAT YOUR TEAM PROVIDES ONCE

Notebook exports - electronic lab notebook entries to search
Scanned batch sheets - older records made searchable with text recognition
Project reports - development and scale-up write-ups
The planned experiment - what the chemist intends to run
Ranked results - similar past work with conditions, outcomes, and a summary

What happens day-to-day: a chemist describes a planned route. The agent searches the notebooks, scanned sheets, and reports, ranks the closest past experiments, and returns the conditions, the outcomes, and a short summary. The chemist starts from what already worked instead of a blank bench.

For the IT person or technical lead who needs to scope the integration.

INQUIRY SOURCES

Electronic lab notebook exports

Scanned or paper batch sheets processed with text recognition

Project and development reports

WORKFLOW ENGINE (N8N)

Chemist queries a plan

Prepare corpus with text recognition

Search the experiment corpus

Rank matches and extract conditions

Any strong match

Summarize what worked

Return results to chemist

CONNECTS TO

ELN (Benchling / IDBS / Signals Notebook)Read: notebook entries and conditions. Write: none

Text recognition (Azure Document Intelligence / AWS Textract)Read: scanned batch sheets. Write: extracted text

Search index (pgvector / Elasticsearch / Pinecone)Read: embeddings of the experiment corpus. Write: none

Notification channel (Slack / Teams / Email)Read: none. Write: ranked results and summary

WITHOUT PMS API

Without a live index the team provides notebook and report exports and scanned sheets in a folder. The agent still searches and summarizes, with the corpus refreshed by hand.

WITH PMS API

With the notebook, text recognition, and a search index in place, the agent searches the full corpus the moment a chemist asks and returns ranked prior work with conditions. Best experience.

For the person who wants to build this agent step by step. Each node below maps to an n8n workflow node.

TRIGGERNode 1

Trigger - a chemist asks

A form or chat entry where the chemist describes the planned reaction or route.

Output: { query_text, reagents[], target }

APINode 2

API - search the corpus

Embed the query and search the index over notebooks, scanned sheets, and reports.

Output: { matches[]: { source, score, excerpt } }

LLMNode 3

LLM - rank and extract conditions

LLM node that ranks the matches and pulls conditions and outcomes from each.

Output: { ranked[]: { source, conditions, outcome, why_similar } }

You are an R&D knowledge assistant. Given the planned experiment {query} and the retrieved records {matches}, rank the records by relevance, and for each pull the key conditions (reagents, solvent, temperature, time) and the outcome (yield, purity, issues). Add one line on why it is similar. Respond ONLY with valid JSON. No explanation.

INPUT

{"query":{"target":"reduce nitro group on intermediate K","reagents":["H2","Pd/C"]},"matches":[{"source":"NB-2019-114","excerpt":"nitro reduction of intermediate K, 5% Pd/C, methanol, 40 C, 4 bar H2, 95% yield, filtration slow"}]}

OUTPUT

{"ranked":[{"source":"NB-2019-114","conditions":"5% Pd/C, methanol, 40 C, 4 bar hydrogen, 4 hours","outcome":"95 percent yield, filtration was slow","why_similar":"Same nitro reduction on intermediate K with the same catalyst"}]}

DECISIONNode 4

Decision - any strong match?

If the top score is above the threshold, summarize. Otherwise return the closest few with a low-confidence note.

Output: { strong_match: boolean }

ACTIONNode 5

Action - return results to the chemist

Post the ranked matches, conditions, outcomes, and a short summary to the chemist.

Output: A ranked result set with conditions and a summary

MVP vs Advanced — Where To Start

MVP — START HERE

Index one project's notebooks and reports
Run text recognition on the scanned sheets
Search by reagents and target
Extract conditions and outcomes per match
Return the top three with a summary
Chemist decides what to run

ADVANCED — ADD LATER

=+ Index the full notebook and report archive
=+ Search by structure or reaction class
=+ Add scale-up records to the corpus
=+ Re-index on a schedule as new work lands
=+ Show a conditions comparison across matches
=+ Flag safety notes attached to past runs

Full Video Walkthrough Coming Soon

Watch the agent surface a three-year-old experiment that matches a planned route.

Everything on this page is free to learn.

When you're ready to build — do it yourself or bring us in.

Book a call to scope this

OTHER AGENTS IN LEASING & MARKETING

Process Safety Route Pre-Screen

Flags reactive-hazard red flags and drafts the pre-experiment safety checklist

Frequently asked questions

R&D Experiment Knowledge Miner is an AI agent that finds similar past experiments and routes so chemists stop repeating old work.

The reaction was run three years ago. The chemist ran it again to find out. The R&D Experiment Knowledge Miner addresses this: finds similar past experiments and routes so chemists stop repeating old work.

No. Your team stays in control. The agent does not design the experiment or decide the route. Your chemist does. The agent removes the manual hunt through old notebooks and repeated work.

This is 2 of 2 agents in R&D & Process Development. → · Explore all 15 departments →