Repository navigation
v0.11.1
The bundle's embedding server now keeps its model loaded, so the first search after a quiet period
no longer waits for the model to load again. The HTTP embedder is reported ready only once its
endpoint answers; when it does not, the detailed /health response gives the reason. Class matching
and MITRE technique search make one embedding call per request instead of one per text. /health
reports the release version. The embedding settings are checked at start, and a blank one means
unset. Modules written in code are no longer reinstalled on every start. A model update tolerates a
class the platform does not have; Dethereal users get this with Dethereal 0.4.13. This is a patch
release: no API, schema or data change. Compared against the previous tag, v0.11.0.
Upgrading: take the new bundle and follow the operator guide's upgrade procedure — back up,
unpack, set PLATFORM_VERSION, ./byodt update. The only line of .env.example that changed is
PLATFORM_VERSION. Take the new bundle's compose.yaml: it sets OLLAMA_KEEP_ALIVE=-1 on the
ollama service. No other .env change is needed. The embedding model stays in memory.
With OLLAMA_KEEP_ALIVE=-1 the model stays resident in the ollama container for as long as it
runs, instead of unloading after 5 minutes idle, so that memory is now held the whole time. A
non-numeric EMBEDDING_SIMILARITY_THRESHOLD now stops the start. A value that is not a number, or
one outside -1 to 1, stops the backend with a message that names the variable; earlier releases
accepted it silently. If the setting was changed from the bundle's 0.40, check it before
upgrading. An empty or blank value now means unset. The first start after the upgrade waits
briefly for the embedding server: the start-up check makes up to three attempts, at most about
half a minute. That start also reinstalls the modules written in code once, because the platform
version is part of their computed hash; later starts skip them while they are unchanged. Dethereal
0.4.13 carries the model-update fix to plugin users. It is published after the platform, and it
also works with platform 0.11.0, since this release changes no API. Do not select an in-process
ONNX embedding profile yet. As with 0.11.0, keep EMBEDDING_MODEL on an HTTP model, such as the
bundle's embeddinggemma.
Changed
- The bundle's embedding server keeps its model loaded. By default Ollama unloads a model after
5 minutes without a request, so the first search after a quiet period waited for the model to load
again. The bundle'scompose.yamlnow setsOLLAMA_KEEP_ALIVEto-1on theollamaservice:
the model stays resident for as long as the server runs instead of unloading after 5 minutes idle,
and the backend's start-up check of the HTTP embedder loads it when the platform starts. The cost
is memory — the embedding model stays in theollamacontainer's memory the whole time. To
upgrade, take the new bundle'scompose.yaml; no.envchange is needed. - Class matching and MITRE technique search embed a request's texts in one call. Class
matching embedded each element's description with its own request to the embedding model, one
after another, and MITRE technique search did the same for each query that matching by ID, name
and description left unanswered, so a request with many elements or queries made as many
embedding calls in a row. Each request now embeds all of those texts in one call, then searches
per element or query; the results are the same. This cuts the round trips to the embedding server,
where every Ollama request carries a fixed cost, and lets the in-process embedder batch the texts.
If the embedding server answers the batch with an error, each text is embedded on its own, in one
attempt without further retries, so a text that cannot be embedded still skips only its own vector
search, with the same logged warning as before. If the server cannot be reached, the request's
texts skip their vector search at once, without a further attempt per text.
Fixed
/healthreports the platform release version. The detailed response'sversionread
0.0.1on every deployment of the platform image. It now carries the release version, such as
0.11.0, orunknownwhen the backend cannot determine it. The production response still carries
only the overall status.EMBEDDING_SIMILARITY_THRESHOLDis checked at startup. The backend accepted any value, so a
setting such ashighorNaNsilently disabled or broke the similarity floor for class matching
and MITRE technique search. It is now checked like the per-corpusEMBEDDING_THRESHOLD_*
settings: a value that is not a number, or one outside -1 to 1, stops the backend from starting
with a message that names the variable. The bundle's default of0.40is unaffected.- An empty similarity floor now means unset. A line such as
EMBEDDING_THRESHOLD_CLASSES=in
.envwas read as a floor of 0. An empty or blankEMBEDDING_THRESHOLD_CLASSES,
EMBEDDING_THRESHOLD_MITRE,EMBEDDING_THRESHOLD_COMPONENTSorEMBEDDING_SIMILARITY_THRESHOLD
now counts as unset, the same as leaving the line out, so the next setting in line or the
embedding profile's value applies. An empty or blankEMBEDDING_THREADSlikewise means the
default of 4 threads; it used to stop the backend from starting. - The HTTP embedder is ready only after its endpoint answers. The backend reported the HTTP
embedder as ready at start without contactingEMBEDDING_URL, so a wrong URL, an unreachable
server or a model with the wrong dimensions showed as ready in/healthand searches then failed
one by one. At start the backend now embeds one short text and expects a vector of the configured
dimensions. If that fails, the embedder isfailed, the detailed/healthresponse gives the
reason, and the backend checks the endpoint again every minute until it answers, then becomes
ready without a restart. The module install at start waits for the check, at most about half a
minute, so the class vectors are embedded on the first start; if the check failed, the modules are
marked partial and a later start embeds them. - Modules written in code are no longer reinstalled on every start. The backend skips installing
an unchanged module by comparing its content hash, but only file modules shipped one, so modules
written in code, such as the Threat Report (dethernety-threat-report) and
dethernety-coverage-tools, were fully reinstalled at every start. The backend now computes a
hash for a module that ships none, from its metadata, the active embedding model and the platform
version. Such a module is reinstalled when its metadata changes (a class, a class field, the
module version), when the embedding model or its dimensions change, or when the platform is
upgraded (once per upgrade), and skipped otherwise. A module that serves its own pre-computed
vectors and ships no hash is still installed on every start.MODULE_LOAD_FORCE=1still
reinstalls every module. - A model update tolerates a class the platform does not have. An update of an existing model,
such as Dethereal'supdate_model, used each element's class id exactly as the model files gave
it. A class the platform did not have, for example because its module is not installed there,
made the update report failure with a message that named neither the element nor the class, even
though the rest of the update had already been written. Update now finds classes as import does:
by id, then by module and class name, so a class with a different id on this platform is still
found. An element whose class the platform does not have keeps its current class and answers, the
update adds a warning that names the element, and the update carries on. The errors that remain,
such as a class binding or an answer write the platform refuses, now name the element and give
the platform's reason. After the push, Dethereal keeps that element's class and answers in the
local model files, so they still show the class you chose. Dethereal users get this fix with
Dethereal 0.4.13.
Documentation
- The embedding configuration references describe the start-up check. The configuration guide
states that an empty or blank similarity floor counts as unset, and that a non-numeric or
out-of-rangeEMBEDDING_SIMILARITY_THRESHOLDstops the backend like the per-corpus floors. It
describes the HTTP embedder's start-up check: up to three attempts and at most 34 seconds,
loadingwhile it runs,failedwith the reason in the detailed/healthresponse, and a new
check every minute until the endpoint answers. It also states that the start-up module install
waits for that check, and that/healthreports the release version orunknown. The BYODt
configuration page explains that the bundle setsOLLAMA_KEEP_ALIVEto-1, so the embedding
model stays loaded and its memory stays held. - The embedding service references follow readiness and batched embedding. The custom resolver
services reference states that the embedding service is enabled only when configured and ready,
documents its pending state, its state values and its wait for the start-up check, and drops the
per-session disable; the pre-computed embeddings specification drops it too. The class and
control resolver architecture describes class matching embedding a request's elements in one
call, and each text on its own when the batch fails. - The module install references describe the content-hash skip. The module management service
reference describes the install phases — the wait for the embedder check, vectors made outside
any transaction, one write transaction per module, then the obsolescence sweep and the
post-commit hooks — and the skip gate, with the hash computed for a module written in code from
its metadata, the active embedding model and the platform version. The module development guide
gains a section on what each kind of module needs for an unchanged install to be skipped. - The model update references describe class resolution. The data-access library's
import/export notes describe the class resolver that import and update now share, and what each
does with a class it cannot find; its operations notes state that update reads the platform's
reason for a refused answer write. The Dethereal sync notes gain a section on class resolution on
push: the warnings, the local files keeping the chosen class, and the errors that remain. The
sync user guide gains a section on classes the platform does not have.