Graph retrieval tied plain vector search at best and never beat it. Measured
across five corpora, the bottleneck was seeding, not the graph: the walk
started from nodes whose embeddings were computed from bare entity names, and
a whole question shares almost nothing with a name like "Quill".
Extraction now embeds each node from "name (type): description" and each
relationship as the fact it asserts ("Alder streams_to Quill: ..."), stored on
a new nullable graph_edges.fact_embedding column that ensure_vector_schema adds
in place. Entity names are canonicalised (case, punctuation, word breaks and a
cautious plural) so "VECTOR_STORE" and "vector stores" land on one node. Extraction calls run
concurrently (GRAPHRAG_EXTRACTION_WORKERS, default 8) while embedding and graph
writes stay serial on the task thread, so ordering and idempotency are
unchanged; that measured 8.4x faster with identical output.
Retrieval gains per-source options, stored under retrieval.graph and read live
at query time:
- seed_strategy: start from matching entities (default) or matching
relationships, which can reach an entity the question never names;
- passage_nodes (on): walk the source's passages alongside entities, with
PageRank damping 0.5 instead of 0.85;
- blend_vector (on): fuse the graph ranking with the source's vector ranking
by reciprocal rank.
The defaults are the measured-best configuration. Through GraphRAGRetriever,
the new seeding moved recall@4 from 0.41 to 0.68 on a multi-hop corpus and
from 0.50 to 1.00 on the docs corpus, and regressed none of the corpora
measured. Existing graphs keep name-only embeddings until rebuilt.
The docs site failed to build: a bare "<= 1" in the prose of the
generated page is parsed by MDX as the start of a JSX tag ("Unexpected
character '=' before name"). Constraints are rendered as code spans now,
where MDX leaves them alone, and a test rejects any bare <, { or } outside
a code span so a future description cannot reintroduce the failure.
Verified with a local next build of the docs site.
Review follow-up. The per-group secret validators normalised a hand-picked
list of API keys, which left other optional credentials and overrides
(OPEN_ROUTER_API_KEY, S3 and Daytona keys, ELASTIC_PASSWORD, the OIDC
trio, connector client ids, MICROSOFT_AUTHORITY, MCP_OAUTH_REDIRECT_URI)
holding the literal "None" or "" a .env file spells "unset" with, so
truthiness checks and fallbacks downstream saw a value. One rule on the
group base replaces those lists: every Optional[str] field maps "", "None"
and whitespace to None and strips real values. Plain str fields are left
alone. The OIDC required-settings check therefore also rejects those
spellings.
EMBEDDINGS_POOLING is Literal["cls", "mean"] with case-insensitive
parsing; its consumer silently ignored anything else.
Bounds added where the consumer rejects or misbehaves on the value:
SCHEDULE_RUN_OUTPUT_RETENTION_DAYS and MESSAGE_EVENTS_RETENTION_DAYS (the
cleanup repositories raise on <= 0), EMBEDDINGS_DELEGATE_TIMEOUT, the
remote-device idle/pairing/invocation TTLs and CELERY_VISIBILITY_TIMEOUT
(> 0), REMOTE_DEVICE_CMD_QUEUE_TTL_SECONDS (> 605, the documented drain
deadline), GRAPHRAG_MAX_CHUNKS_FOR_EXTRACTION (>= 0; negative would slice
the pending list from the end).
The generated reference now renders generic type arguments
(dict[str, int] rather than dict).
About 85 call sites read a setting as getattr(settings, "NAME", fallback),
each carrying its own copy of the default. Every one of those names is a
field with a default on the model, so the fallback could never apply to
the real settings object; it only masked drift. Two had drifted:
- OPENAI_PROMPT_CACHE_KEY defaults to True on the model but the reader
fell back to False, and two test stubs relied on that.
- SharePoint's MICROSOFT_AUTHORITY fallback to
https://login.microsoftonline.com/<tenant> never fired, because the
attribute always exists (as None), so MSAL got authority=None. The
connector now derives the tenant authority when the setting is unset,
as its test always assumed.
Four places read EMBEDDINGS_KEY straight from os.environ, skipping the
"None"/"" normalisation the model applies; they read the setting now.
Test stubs that replaced a module's settings with a SimpleNamespace list
every setting the code under test reads.
SAGEMAKER_REGION, SAGEMAKER_ACCESS_KEY and SAGEMAKER_SECRET_KEY survive
only as a fallback for the S3_* credentials. They carry
Field(deprecated=...) now, so any read emits a DeprecationWarning naming
the replacement and the generated reference shows the notice. The S3
store is the one sanctioned reader; it silences that warning locally
because it already logs its own operator-facing one when the fallback
is actually used.
DEFAULT_MAX_HISTORY was referenced nowhere. RETRIEVERS_ENABLED was read by
no code at all, while two docs pages described it as an enforced
allow-list; both the setting and those claims are removed.
The "AUTH_TYPE=oidc requires OIDC_ISSUER, OIDC_CLIENT_ID and
OIDC_FRONTEND_URL" check lived in app.py, so it only ran when the Flask
app was imported; a worker or script with the same misconfiguration
started fine. It is now a model validator on the auth group and runs
wherever Settings is loaded, with the same message.
DEPLOYMENT_TYPE, which app.py read straight from the environment to
decide whether a missing JWT_SECRET_KEY is fatal, is a documented
setting on the server group now, so it shows up in the reference like
every other variable the app reads.
Enum-like settings whose allowed values were only listed in a comment are
now Literal types, so a typo fails at startup with a message naming the
allowed values instead of falling through to a default with a warning
(or, for VECTOR_STORE, failing on first use):
AUTH_TYPE, VECTOR_STORE, STORAGE_TYPE, URL_STRATEGY, OCR_BACKEND,
OCR_ENGINE, SANDBOX_BACKEND, DOC_PARSER_ENGINE, TTS_PROVIDER, STT_PROVIDER
Each keeps a before-validator that strips and lower-cases the value, since
the registries that consume them already lower-cased at the use site, and
AUTH_TYPE maps the "None"/"none"/"" spellings a .env file carries to None
(it was the string "None" before, which only worked because nothing
compared against it). An empty TTS/STT provider still means "off".
LLM_PROVIDER stays a plain str because providers are plugin-extensible.
Containers are typed (dict[str, int], list[str], dict[str, Any]) instead
of bare dict/list, six fields that were Optional with a non-None default
are plain, and integer settings whose description already states a range
carry it as a constraint (ge=0 for "0 disables", ge=1 for counts that
cannot be zero, 0 < threshold <= 1).
The hand-maintained settings page documented 95 of 258 settings and
.env-template 42, and both drifted as fields were added. The field
descriptions now live on the model, so the reference is rendered from it:
python -m docsgpt.core.settings.reference --write
writes docs/content/Deploying/Settings-Reference.mdx, one section per
settings group with each field's type, default, constraints, aliases and
description. --check reports a stale page, and tests/core/test_settings.py
fails when the checked-in page no longer matches the definitions, so a
new setting cannot land undocumented.
test_settings.py also pins the composition contract: every group field is
a flat Settings attribute, no field is defined twice, every field has a
description, and the secret-normalising validator of every group is
applied (the case that a shared method name would silently drop).
The App Configuration page points at the reference instead of at
settings.py, and the reference is listed in the Deploying navigation.
Alex