✅ every quote verified🔒 keys never stored🧠 open-weight friendly🌐 runs in your browser📚 open data
The ideaWhat is PathMap Alchemy?
Science publishes millions of papers, and no one can read them all. Sometimes idea A is linked to a middle idea B in one set of papers, and B is linked to C in a completely different set, but nobody has connected A to C. This is called literature-based discovery (the Swanson method), and it has produced real medical hypotheses.
PathMap Alchemy automates that search. It was built for pathology and nutrition questions, such as which foods, compounds or mechanisms might connect to a disease, but it works on any two biomedical concepts.
🔗
Finds the bridge
Looks for the shared middle term that links two ideas that never appear together.
✅
Proves every claim
Quotes are exact sentences copied from PubMed abstracts and matched by code, then audited for drift and invention.
🧪
Adds the chemistry
Pulls molecule data from PubChem and its PUG View facts, plus taxonomy and PubMed context.
🧭
Checks novelty
Counts how often A, the bridge and C already appear together, so you can tell what is new.
🕸️
Builds a library
Saves every run on your device and links them in a knowledge graph you can export.
♾️
Explores on its own
Autonomous mode chains follow-up questions automatically and never repeats a pair.
Under the hoodHow a run works
SearchThe AI writes careful PubMed queries for each concept, favouring the food, plant or compound meaning of ambiguous words.
ReadAbstracts are retrieved from PubMed and cleaned. Small models get a sampled set so their context window never overflows.
HypothesizeThe AI proposes one bridge and an A to B to C hypothesis, citing evidence only by quote ID.
VerifyCode matches each quote to the real abstract, then a strict audit checks for drift. Failures trigger another attempt, and the result gets a Grounding Score.
ReportYou get a knowledge graph, validated quotes with emojis, molecule cards, figures, a novelty probe, and a PDF.
⚠️ Results are hypotheses produced by an AI reading abstracts. They are not medical advice, are not peer reviewed, and should be checked against the original papers.
Get goingWhat you need
Alchemy runs entirely in your browser and works with the AI of your choice:
☁️
Gemini (free key)
The easiest start. Get a free key from Google AI Studio and pick a Gemma or Gemini model.
💻
Your own computer
Use open-weight models with Ollama or LM Studio. Private, offline-capable, no limits.
🔀
OpenRouter
One key, many models, including paid ones.
🔒 Your API key stays in memory for the open tab only. It is never saved, never put in traces, and is sent only to the provider you chose.
PathMap Alchemy searches PubMed for the hidden link between two ideas and checks every quote against the real abstracts. It runs entirely in your browser. This guide covers every button and what to use it for.
Quick start
1
Connect a model. Tap the gear, choose Gemini, Localhost or OpenRouter, paste a key if needed, and press Save. The top pill turns green when it works.
2
Pick a mode (Discovery, Literature review or Autonomous) and type your concepts. A question is optional.
3
Press the big green button and watch the log. A run usually takes a few minutes.
4
Read the report, then download it as PDF, check it with Dig Deeper, or run a suggested follow-up.
5
Find it again later in My Library. Every finished run is saved automatically on this device.
Top menu
🌐
Language
WhatChanges the page language and the language the AI writes the report in (English, Español, Français, Deutsch).
WhenYou would rather read results in another language.
WhyThe quotes stay in the original English of the abstracts so they remain verifiable.
●
Connection pill
WhatShows the model and provider when connected, or "Not connected".
WhenCheck it before starting a run.
WhyIf it is grey, runs cannot start. Tap the gear to connect.
+ −
Zoom
WhatThe + and − buttons make the whole page larger or smaller (80% to 200%). Tap the percentage to reset to the default, 130%.
WhenText is hard to read on a phone or small screen.
WhyYour choice is remembered on this device. PDFs always print at normal size.
Guide
WhatOpens this help window. Close it with the ✕ button, the Esc key, or by tapping outside.
WhenAny time you forget what a button does.
WhyEverything in one place, with the icons you see in the app.
Download (button and icon)
WhatSaves the entire app as one HTML file (pathmap-alchemy.html).
WhenYou want an offline copy, or to run it with a local model like Ollama.
WhyNo server needed. The file is the whole app.
My Library
WhatOpens every run you have saved.
WhenYou want to reopen, compare, export or organise past results.
WhyYour discoveries build up into a searchable, linked collection.
? About
WhatExplains what the app is, how it handles your data, and lets you reset your sharing choice.
WhenYou want the privacy details or to re-download the app.
WhyTransparency about what stays on your device.
Settings (gear)
WhatChoose the AI provider, paste a key, pick the model, and switch the Gemma 16-abstract limit on or off.
WhenFirst use, or when you change model or provider.
WhyAlchemy works with free Gemini keys and with open-weight models on your own computer.
Choose a mode
🔗
Discovery
WhatFinds ONE hidden A → B → C link: two concepts that never appear together, joined by a shared middle term (the bridge). Every claim is backed by exact quotes.
WhenYou have two ideas that might be connected in a way nobody wrote down, for example a food and a disease.
WhyThis is the Swanson literature-based discovery method. Alchemy is allowed to report that nothing is there.
📚
Literature review
WhatSummarises what PubMed says about one topic (a second concept is optional): themes, contradictions, evidence gaps, and suggested studies.
WhenYou want the state of the evidence rather than a new idea.
WhyFast, quote-backed grounding in a topic before you go deeper.
♾️
Autonomous
WhatRuns your seed as a Discovery, then automatically picks a random unexplored follow-up pair from all suggestions so far, runs it, and repeats. It never repeats a pair (including pairs already in your library).
WhenYou want to explore a whole territory hands-off, for example overnight.
WhyIt stops when there are no new pairs left, after 3 failures in a row, or when you press Cancel run. Each run is saved, and at the end you can open the whole session as a book.
✨
Suggest for me
WhatFetches a random food and health-condition pair from pathmap.org and fills Concept A and Concept B for you. You still press the run button yourself.
WhenYou are not sure what to explore, or want a fresh idea.
WhyQuick way to try the app. Press it again for a different pair, or edit the fields before running.
💬
Question / cue (optional)
WhatA sentence that steers the AI, for example "Could steam from cooking reach the brain?"
WhenYou have a specific question in mind.
WhyThe report includes an "Answer to your question" section that answers it directly from the evidence.
While it runs
🖥️
Thoughts log and timer
WhatA live log of every step: query writing, PubMed search, quote checking, PubChem lookups, audits.
WhenWatch it if you are curious, or ignore it.
WhyNothing is hidden. If a step fails the log says why.
⛔
Cancel run
WhatStops the current run (or the whole autonomous session).
WhenIt is taking too long or you typed the wrong concepts.
WhyFinished runs stay saved. Only the run in progress is lost.
🔁
Retries and audits
WhatYou may see "attempt 2" or a "Restart". The app re-checks quotes and claims and tries again until they hold up.
WhenIt happens automatically.
WhyThis is how it avoids invented citations. A few retries are normal.
Reading the report
🕸️
Knowledge graph
WhatA map at the top of the report linking this run to its concepts, bridge, molecules, MeSH tags and cited papers.
WhenLook at it first for the big picture.
WhyShows at a glance what connects to what. It prints in the PDF.
✅
Veridicality banner
WhatA green or red box with a Grounding Score from the strict audit.
WhenCheck it before trusting anything.
WhyRed means treat the claims as unverified. Use "Run again".
🍗
Validated quotes
WhatExact sentences copied from PubMed abstracts and matched by code, not typed by the AI. Each has its own emoji.
WhenUse them to verify any claim yourself.
WhyThe same emoji marks the same quote everywhere in the report, so you can trace claims back easily.
🔎
Clickable PMIDs
WhatEvery PMID opens that paper on PubMed in a new tab.
WhenYou want to read the original abstract.
WhyNever trust a summary without checking the source.
🧭
Novelty probe
WhatPubMed counts for A + bridge, bridge + C, A + C, and all three together.
WhenJudging whether the link is really new.
WhyThe first two should be above 0 (that is the bridge). If A + C is above 0 the link may already be published.
🧪
Molecule cards
WhatChemicals, nutrients and toxins in the discovery, with PubChem data (CID link), PUG View facts, taxonomy and PubMed context counts.
WhenYou want the chemistry behind the idea.
WhyProperties like size, charge and lipophilicity can support or weaken a hypothesis. Plants and proteins may show "no PubChem record", which is normal.
💡
Follow-up ideas
WhatAI-suggested next searches with a "Run this experiment" button.
WhenYou want to keep exploring from this result.
WhyDiscovery chains naturally: one result suggests the next question.
📐
Figures and tag cloud
WhatDiagrams drawn by code from the verified data, with Download SVG buttons, plus a cloud of key terms.
WhenYou need figures for a talk or paper.
WhyIdentical every time because no AI draws them.
Buttons above the report
⬅
New synthesis
WhatReturns to the home screen.
WhenStart another search.
WhyYour finished run is already saved.
📄
Download Synthesis as PDF
WhatOpens your browser print dialog. Choose "Save as PDF".
WhenYou want to share or archive the report.
WhyIncludes the graph, quotes and figures.
🔍
Dig Deeper with Google / ChatGPT
WhatOpens Google or ChatGPT with a ready-made question about the hypothesis.
WhenYou want an independent second opinion on whether it is real and new.
WhyAlchemy generates hypotheses, not facts. A separate check is good practice.
🧾
Download full trace
WhatA complete JSONL record of the run: every prompt, AI reply, PubMed and PubChem request, build metadata and build log.
WhenYou need to audit or debug a run, or keep proof of provenance.
WhyFull reproducibility. Your API key is never included.
WhenYou want to analyse results in a spreadsheet, notebook or database.
WhyOpen data you can reuse anywhere.
Download App (.html)
WhatSaves a copy of the app itself.
WhenYou want a portable copy.
WhySame as the header download button.
My Library
Every finished run is saved automatically in this browser (up to 200). Open it with the My Library button.
🗂️
Projects
WhatCreate a project, add selected runs to it, or filter the list by project.
WhenYou are working on several topics at once.
WhyKeeps related runs together and makes books and graphs focused.
☑️
Select, delete, export
WhatSelect all or some runs, then export them as JSONL or delete them. Clear library removes everything.
WhenBackups, sharing datasets, tidying up.
WhyBrowser storage can be cleared by your browser. Export anything you care about.
📖
Open as book
WhatCombines the selected runs into one document with a table of contents and a project knowledge graph.
WhenYou want one PDF for a whole project.
WhyTurns many runs into a single readable report.
🕸️
Graph view and Export Cypher
WhatAn interactive graph of runs, concepts, molecules, MeSH tags and papers, with a shortest-path finder. Cypher export is for Neo4j 5.
WhenFind shared molecules or surprising paths between runs.
WhyReveals connections no single run shows.
⬆
Import
WhatLoad a saved trace, JSON or dataset file back into your library.
WhenMoving to a new device or restoring a backup.
WhyYour work is portable.
🏷️
Molecule and MeSH chips
WhatClick one to list every run that shares it.
WhenLooking for everything you have found about one compound or topic.
WhyFast cross-referencing.
Settings details
☁️
Gemini API
WhatPaste a free key from Google AI Studio and pick a model (Gemma 4 is open-weight).
WhenEasiest way to start.
WhyFree tier is enough for many runs.
💻
Localhost
WhatPoint to Ollama or LM Studio on your own computer. Start Ollama with OLLAMA_ORIGINS=* so the browser may call it.
WhenYou want privacy, no limits, or offline use.
WhyNothing leaves your machine except PubMed and PubChem lookups.
🔀
OpenRouter
WhatPaste an OpenRouter key and a model name.
WhenYou want a different or paid model.
WhyWide choice of models with one key.
🔒
Your API key
WhatThe key stays in memory for this tab only. It is never saved, shared or put in traces, and it is sent only to the provider you chose.
WhenReloading the page clears it, so paste it again.
WhySafe on shared computers.
✂️
Gemma 16-abstract limit
WhatSwitch that limits Gemma models to a random sample of 16 abstracts per prompt.
WhenKeep it ON for small context windows. Turn it OFF for localhost or paid models with large context.
WhyMore abstracts can give richer results when the model can handle them.
Sharing (optional)
🌍
Share with the global library
WhatBefore a run you are asked whether to share the finished result with the PathMap research library. Choose Keep private to share nothing.
WhenIf you want your discoveries to contribute to open research.
WhyShared data: concepts, write-up, quotes, molecules, model name, app version and the run trace. Never your API key. Use "Learn more" in the dialog for details, and reset your choice from the ? About window.
Tips and troubleshooting
🍗
Ambiguous words
WhatWords with two meanings are read as food first, so "turkey" means the meat, not the country.
WhenAdd detail if you mean something else, such as "Turkish population".
WhyGets you better search results.
🎯
Better concepts
WhatSpecific beats general: name a mechanism, compound or condition rather than a whole field.
WhenResults feel vague or "no overlapping evidence".
WhySpecific terms retrieve more relevant abstracts.
⚠️
Common messages
What"Could not reach the server" usually means a local endpoint or CORS problem. "Rate limit" means wait a minute or switch model. "Domains differ from your input" means the data supported a different link than you asked for.
WhenSomething stops or looks odd.
WhyAll results are hypotheses from AI reading abstracts. They are not medical advice and are not peer reviewed.
Open dataSharing, ARKs and open datasets
Before each run the app asks whether you want to share the finished result. Sharing is optional, and "Keep private" sends nothing.
🧼
Sanitized first
Before anything is sent, API keys, local or private network addresses, browser details and email addresses are removed.
🔎
Reviewed
Not every submission is added or accepted, but every valid submission we receive is processed for publication.
🔖
Permanent ID
Accepted results receive an ARK, a permanent identifier that can serve as a dated, citable record of the idea.
📦
Open datasets
Accepted results are published as open JSONL datasets, including on Zenodo, for research and for training open-weight models.
What is shared: your concepts and cue exactly as typed (do not put private notes in them), the write-up, verified quotes with PMIDs, molecule data, the model name, the app version, and the run trace. Never shared: your API key, your local address, or other runs in your library.