October at Inflexa: analyses that survive a crash
Work that picks up where it stopped, citations checked against real lookups, report figures drawn the way each field draws them, and better support for the model you bring. Here is what changed and why.
Six changes stand out. Here is each one, and why we made it.
- 01Work survives a failureA crash or an outage no longer costs you the turn.
- 02Citations from real lookupsA PubMed ID no lookup returned is refused.
- 03The figure your field expectsReport charts are drawn from the run's tables.
- 04Your own model, with a warningA context window that is too small is flagged.
- 05Approvals, patents, the webNew tools for regulatory and web research.
- 06A bill of materialsOne command lists what you installed.
An analysis now survives a crash or an outage.
A chat turn used to live inside the request that started it, so a crash or a restart lost it. Now each model call and each tool call is a checkpoint. After a crash, the turn continues from the last one, and nothing that finished runs again. Sub-agents work the same way, and two sub-agents of one reply now run at the same time.
When a model provider goes down, the work retries after a delay and then pauses. A sign-in that has expired pauses it right away. You see the cause, then resume from the same place or stop. Runs and data profiles pause the same way.
Sandboxes follow the same idea. Each one is now held by a lease that outlives the step using it. If a machine dies it is replaced, and a command that already finished returns its first result.
- model callsaved
- lookupsaved
- sub-agentsaved
- model callsaved
The model provider is not answering. Retries ran out.
r resume · s stop
A resume picks up from the last checkpoint. The saved calls do not run again.
A cited paper has to come from a lookup.
“Are these references real?” comes up whenever we talk with scientists about AI in analysis. When Inflexa writes up the findings of a run, it now refuses a PubMed ID that no lookup in that run returned, so the agent has to look the paper up or leave it out.
That shows the paper exists and was looked up. Whether it supports the finding is still a person's call.
The record itself also holds more. Each file a step reads now enters its lineage, and each script the agent runs is recorded with the hash of the script, of the tables it read, and of the files it derived.
How provenance works- PMID 28129391
- PMID 30143610
- PMID 31415926
pmid "31415926" is in no lookup result of this run
- readcounts.tsv#7ac2e01b…
- readprotocol.pdf#c41d9f36…
- scriptrank.py#0fb83d72…
- derivedtop_hits.tsv#e96a1c05…
Report charts draw the figure your field expects.
Reports used to place a run's PNGs on the page. Now a report chart is drawn from the run's table, as the standard figure of its field: a volcano for differential expression, a UMAP for single cell, a Manhattan plot for GWAS, a Kaplan-Meier curve for survival.
We measured which plots papers actually use, read the field's guidance on how to draw each one, and tested every figure on public data. On a dense chart the axes and text stay vector, over a 300 DPI layer of points.
A chart you already have takes the new look the next time you preview it.
How reporting worksBring your own model, and know if its window is too small.
The conversation agent needs room to work. A chat model with a context window under 200,000 tokens used to overflow with no error. One small model on Ollama answered with no tool calls at all.
The CLI now checks the window when you switch models and warns you when it is too small. A smaller window also compacts earlier, at 80% of its size. Ollama reports its window on its own. For any other endpoint, you can set contextWindow in your config.
The warning is there because a small model can still struggle, so treat it as a real signal. Connections to OpenAI also move to the Responses protocol, and a picture a tool returns, such as a rendered report page, now reaches the model.
The setup commandlocal-model:8b
local-model:8b has a 32,768-token context window. Inflexa needs at least 200,000 tokens and a model with good tool calling.
The agent can search approvals, patents and the web.
Two new tools read FDA and EMA drug approvals, down to the text of labels and reviews, and search patents. Both go through Amass with your own key, so each search spends your Amass credits.
Two more reach the open web. web_search runs on a SearXNG service that the CLI starts for you, and web_fetch pulls the readable text of a page. Both refuse private network addresses by default. Searches go out to public search engines, the same way database lookups go out to public sources.
- search_regulatory_approvalsFDA and EMA labels, reviews, SmPCs and EPARsyour Amass key
- search_patentsPatent searchyour Amass key
- web_searchA SearXNG service on your machineno key
- web_fetchThe readable text of a pageno key
A bill of materials for what you install.
A security review usually starts with the question of what is inside. inflexa sbom prints one CycloneDX 1.6 document for your installation: the CLI binary, the two sandbox images, and the package store.
Each published build now writes its own bill of materials and attests it, and a build fails if any component has no resolved license. The store's list covers the published catalog, so a package you add with inflexa store add is yours to track.
inflexa sbom > inflexa.cdx.json- CLI binary
- sandbox image
- provisioner image
- package store
A local server, and a lot for Inflexa Cloud.
In Inflexa Cloud, one place now stops, resumes and holds the active work of an analysis, and a long run no longer fails when its eight-hour credential runs out. A banner tells you when a model provider reports a problem. Organizations get an activity log with filters and CSV export, and you can rate each reply.
In the CLI, the terminal app now talks to one local server, so you can open your analyses in more than one terminal. If that server crashes, the chat offers to start it again. You can abort a run from the terminal, and a step that needs reference data you have not installed refuses to launch and names each dataset. Scripts can also read a table or a PDF inside a compressed file.
The CLI can now connect to the Knowledge Plane, which is in preview on commercial plans.
- Differential expression runrunningstop
- Chat turnpausedresumestop
- Data profilewaits to retryholdstop
Get the latest release
Already installed? Run inflexa upgrade to bring your install up to date.