The chunker caches only a model's tokenizer.json in the embeddings cache.
FastEmbed counts any cached snapshot as the model, so it never downloaded
the ONNX graph and every load failed with NO_SUCHFILE. The loader now
checks the files FastEmbed needs and fetches them first when they are
missing; offline it leaves them for FastEmbed to report.
Celery records the executing task on the thread that runs it, so a thread that
task starts sees none. The embeddings client and read_document both decided
"am I in a worker?" from that alone, and from any other thread took the
web-process branch: dispatch to the worker they were running in and block on
the result. Celery refuses that get() ("Never call result.get() within a
task!"), so the embed failed and latched the 30s dispatch cooldown for every
caller after it; with joins allowed, read_document would instead wait on a
parsing queue only its own busy process serves.
Threads inside tasks are not hypothetical: per-source retrieval fans out to a
pool, so a scheduled or webhook agent searching several sources embedded from
pool threads. Graph extraction did too, which failed every chunk of a build.
in_worker() in celery_init answers for the whole process. Celery's
task_join_will_block is process-wide and set for every blocking pool (prefork,
solo, threads) -- exactly the condition under which dispatch-and-wait goes
wrong; eventlet/gevent leave it unset, so the task's own thread still counts
through current_worker_task. Verified with real workers on each blocking pool:
from a thread a task started, the old check dispatched and hit the error, the
new one embedded locally.
- Ship tiktoken's cl100k_base inside the package and build the encoding
from it, so token counting never downloads anything.
- Default EMBEDDINGS_CACHE_DIR to <data home>/models instead of FastEmbed's
temp dir, and read tokenizer.json and repo metadata from that cache, so
a model downloads once and survives reboots.
- TTS_PROVIDER=none and STT_PROVIDER=none switch the speech features off:
the endpoints return 404, audio files fail to ingest with a clear
message, /api/config reports tts_available/stt_available, and the UI
hides the Speak and microphone buttons.
- Drop the Google Fonts Roboto import from the web UI.
- prefetch-models fills the cache the app reads; verify-offline checks the
packaged encoding.
- Docs: new Air-Gapped Deployment guide, settings and cache notes.
The backend import package is now docsgpt, the name it will carry on PyPI;
application was far too generic to install into anyone's site-packages.
git mv plus a mechanical rewrite of every import, dotted string and path
reference: 734 Python files, the compose files, Dockerfile, workflows, docs,
setup scripts, devcontainer, k8s manifests, vscode config, pytest and coverage
config, .gitignore. Behaviour is unchanged.
Kept for one release:
- A top-level application package whose meta-path finder resolves
application.x.y to the already-imported docsgpt.x.y object, so old imports
and entry points (celery -A application.app.celery,
uvicorn application.asgi:asgi_app) keep working with a FutureWarning.
- Celery registers every application.* task name as an alias of its
docsgpt.* task on start-up, so messages queued by the previous release still
run. The redbeat key prefix moves to redbeat:docsgpt:v2: so schedule entries
the previous release wrote are left unread instead of firing twice.
The backend image builds from the repository root (docker build -f
docsgpt/Dockerfile .) so it can ship the alias package; a root .dockerignore
allow-lists docsgpt/ and application/ and keeps caches, local data, .env
files, the sample index files and the Dockerfile out. Compose and the image
workflows point at the new context.
Five defects from a review of the embeddings work, four of them silent.
- Write local files atomically. `LocalStorage.save_file` streamed straight onto
the destination, so an interrupted write left a truncated file. `reembed`
rewrites every index it touches, and a half-written `index.faiss` loads at
neither the old width nor the new one -- the source was unrecoverable, with
no backup and no temp file left behind. Bytes now land beside the destination
and move into place with `os.replace`. S3 was already safe (single PUT).
- Read pgvector chunks a page at a time. `reembed_pgvector` materialised every
`(id, text)` row for a source before embedding -- ~1.6 GB at 200k chunks and
several times that for non-Latin scripts, with the `PGresult` held alongside
until the cursor closed. Inside the shipped 4Gi limit, while also holding the
model, that is an OOMKill -- which is exactly the SIGKILL the point above
turned into a destroyed index. It now walks the source by keyset.
- Bound the first wave of delegated embeds. The failure cooldown is only latched
once the first `get()` returns, so every request already in flight paid the
full EMBEDDINGS_DELEGATE_TIMEOUT: measured 64 threads all timing out together,
and at the shipped 60s across a 96-thread WSGI pool that is an API serving
nothing at all, health checks included. One caller now probes while the rest
fail fast; after a single success the gate leaves the path entirely.
- Ship EMBEDDINGS_NAME commented in .env-template. The comment directly above it
says to leave it commented when upgrading, and the line shipped set. Any value
reaching `.env` lands in `model_fields_set`, which makes `resolve_embeddings_pin`
bail -- so a template-derived `.env` disabled the legacy pin outright and
repointed a populated index at a different 768-dim model, where no width check
fires. The pin already picks granite for a fresh install and mpnet for an
existing one, so nothing needs to be set by hand.
- Stamp `sources.model` on wiki sources. They were created with the column NULL
and then embedded like any other source, and the boot check reads NULL as
"pre-dates the column, therefore the legacy model" -- reporting a correctly
embedded source as stale on every startup of every process. Stamped at
creation, and again on each page re-embed so existing rows heal.
The two docs that promised the FAISS index survives a failed run said so of the
embed only; both now describe the write, and upgrading.mdx says to stop ingest
for the duration.
Query embedding moved to the Celery worker, but nothing that ships was
updated to consume the queue it dispatches to.
- Add `embeddings` to every worker `-Q` list (compose x3, k8s, devcontainer,
sandbox README). Without it a search blocked for EMBEDDINGS_DELEGATE_TIMEOUT
and then answered with no retrieved context, because classic_rag swallows the
dispatch error and skips the source -- bad answers, not an error.
- Skip the task_postrun heap reclaim for the embed task. The full gc.collect()
was written for docling/torch parses; on a worker holding the ONNX model it
measured ~86ms against ~8ms for the embed itself, a 9x slowdown of the round
trip for a task that allocates a few kilobytes.
- Resolve the installation pin in the re-embed script. It never imports
application.app, so an install pinned in app_metadata with no EMBEDDINGS_NAME
set -- every stock k8s deployment, whose manifests carry no embedding config
-- would rewrite its whole index with the legacy default and stamp
sources.model to match, then be told by the boot warning to run it again.
- Fail fast for 30s after a failed dispatch. fanout.embed_questions falls back
to letting each store embed its own query, so one dead-worker retrieval paid
the timeout once in the fan-out and again per source.
- Forget the task result. Nothing reads it back: the key is per-dispatch UUID,
not content-addressed, so a repeated query mints another. Left alone every
search leaked ~17KB for result_expires (7 days) into the Redis the broker
shares -- on the bundled k8s manifest (1Gi, no maxmemory policy) that is an
OOMKill that takes the broker with it.
- Release the model ensure_vector_schema loads to read the width of an
unregistered model, in a process that delegates and would never call it.
The width still comes from the model, not the table, so the mismatch check
the hook exists for keeps working.
- Correct the docs that said otherwise: embeddings.md claimed the standard
deployment worked unchanged, upgrading.mdx said no action was needed, and
the settings table listed none of the three delegation settings.
The API embeds every query it serves, so it held its own copy of the model:
~890 MB it never needed. EMBEDDINGS_DELEGATE_TO_WORKER (on by default) sends
the text to the Celery worker instead and gets the vector back, taking an API
process from 1176 MB to 285 MB with no ONNX Runtime imported at all. The client
embeds locally when it finds itself inside a worker task, so the worker never
dispatches to itself -- the same self-deadlock DOCUMENT_PARSE_QUEUE avoids on
the parsing side. EMBEDDINGS_BASE_URL still wins over it, and remains the right
answer for production.
ensure_vector_schema was constructing the embeddings instance purely to read
.dimension off it, loading several hundred MB of ONNX into every API and worker
process at import. For a model the registry describes that is a lookup; only an
unregistered name now falls back to loading.
EMBEDDINGS_BATCH_SIZE was sizing two unrelated things: chunks per store
transaction (and per remote embed request) and documents per ONNX forward pass.
Each pass pads every input up to its longest, and that waste grows with the
square of chunk length, so at the 1250-token default a batch of 32 peaked at
6.6 GB and took 326s where a batch of 1 peaked at 2.9 GB and took 90s. The
forward pass is now sized by EMBEDDINGS_MODEL_BATCH_SIZE, defaulting to 1;
storage and remote batching are unchanged at 32.
reembed embeds in-process: a batch job that walks the whole index should not
round-trip every chunk through a broker, and loading the model there reports a
real failure instead of timing out against an empty queue.
Also drops the mpnet zip download from the docs and the devcontainer, which
pointed at a SentenceTransformers export with no ONNX graph and had been inert
since the FastEmbed swap; corrects the claim that any sentence-transformers
model works; and settles the Configuring/Settings pages on what the registry
and the repository metadata actually decide.
Follow-up review pass over the embeddings branch.
- Fold an oversized header back into the body, and drop header duplication
when it would leave under a quarter of the chunk budget. A header at or
over max_tokens collapsed the body budget to one token, so a document
became one chunk per body token, each still over the cap: a 95 KB file
produced 20k chunks of 2563 tokens against a 1250 cap. Also clamp
max_tokens to at least 1, as the strategy chunkers already do.
- Emit a header-only document as its own chunk. With no body piece to
attach it to, splitting returned nothing and the document was dropped
from the index with no error and no log line.
- Skip add_custom_model for a repository FastEmbed already ships. It
rejects a name it knows, so configuring any of its ~30 built-ins
(MiniLM, bge, e5, gte, ...) failed every embed call and every query.
- Decide "the user chose this model" by comparing against the field
default rather than model_fields_set, which is true for anything read
from .env. Every setup script has always written EMBEDDINGS_NAME, so an
upgraded remote-embeddings install inherited mpnet's 384-token window
and silently clipped ~80% off every chunk.
- Cut tiktoken splits at character offsets instead of decoding each token
window. A multi-byte character straddling a boundary decoded to U+FFFD
on both sides, destroying one character at roughly one boundary in five
on CJK text -- including at the default max_tokens of 2000.
- Let the re-embed script open a FAISS index whose width does not match
the configured model. That mismatch is the main reason to run it, and
the error recommending the script was raised by the script itself, so
the advice failed on every source.
- Re-embed graph_nodes.name_embedding when GraphRAG is enabled. Those
vectors seed every traversal and share the chunk vectors' width, so a
same-width model swap left the graph retrieving from the old space with
nothing to report it.
- Prefetch the models before copying the application source, so editing
any file no longer re-downloads ~780 MB of artifacts on every build.
- Mirror the setup.sh embedding menu into setup.ps1: granite default,
legacy mpnet as an explicit option, and both engine flows updated.
Windows users were otherwise stranded on mpnet with no granite path.
- Drop the unused EmbeddingsWrapper.tokenizer property.
Follow-up to the embeddings work, from a review pass over the branch.
- Route the OpenAI/Azure key handling through the model registry instead of
matching the canonical name literally, so the `text-embedding-ada-002`
alias the registry now accepts also reaches the Azure deployment name
rather than failing every embed with DeploymentNotFound.
- Fall back to a default width where the embeddings model reports no
dimension. A model outside the registry returns None rather than no
attribute, so `getattr` with a default did not catch it and the width
reached the DDL as `vector(None)` / `list_size=None`.
- Point HF_HUB_CACHE at the prefetch directory. Chunking loads the tokenizer
through `tokenizers`, which reads the hub cache, so a fresh container
fetched over the network on first ingest and an offline one silently fell
back to cl100k.
- Charge a token that collapses a long unbroken run by its character span.
WordPiece emits one [UNK] for any word over its character limit, which made
base64 and minified content count as near-zero tokens, so nothing split it
and oversized chunks reached the embedding server.
- Preserve chunk ids and honour --batch-size when rebuilding a FAISS index.
Fresh uuids orphaned GraphRAG's graph_node_chunks rows, and the whole index
went out in a single embed call on remote servers.
- Document that granite runs an int8-quantised graph, and scope the
SentenceTransformer parity claim to mpnet's fp32 graph, which is where it
was measured.
- Correct the embeddings docs: a matching dimension is not a matching model,
so a same-width swap raises nothing and silently degrades retrieval.
Bumps the Python backend dependency set, holding back the ones that are
resolver-blocked or that regress behaviour this repo depends on.
Notable upgrades:
cryptography 46.0.7 -> 50.0.0 (requires msal 1.37.0, which relaxes its cap)
protobuf 6.33.6 -> 7.35.1 (floats opentelemetry-* to 1.44.0)
openai 2.32.0 -> 2.53.0, anthropic 0.88.0 -> 0.121.0
google-genai 1.73.1 -> 2.17.0 (the 2.0 break is scoped to the Interactions
API, which this repo does not use)
fastmcp 3.2.4 -> 3.4.6, gunicorn 25.3.0 -> 26.0.0
starlette 1.0.0 -> 1.6.0, uvicorn 0.42.0 -> 0.52.1
sentence-transformers 5.3.0 -> 5.7.0, numpy 2.4.4 -> 2.5.1
faiss-cpu 1.13.2 -> 1.15.0, pillow -> 12.3.0 (pinned; security release)
pypdf 6.9.2 -> 6.15.0, lxml 6.0.2 -> 6.1.1 (CVE-2026-41066)
Code changes needed by the bumps:
- openai >= 2.53 rejects a falsy api_key at construction. Keyless
OpenAI-compatible backends (Ollama, llama.cpp, vLLM) legitimately have
none, and pydantic-settings yields "" for a bare `API_KEY=` in .env, so
both call sites now fall back to a placeholder.
- Flask >= 3.1.2 tears a stream_with_context request down twice, and a
ContextVar token may only be reset once. The log-context teardown hook
is now idempotent.
- sentence-transformers renamed get_sentence_embedding_dimension to
get_embedding_dimension in 5.4; use the new name with a fallback.
Held back deliberately:
torch 2.11.0 - 2.13.0 drags torchvision 0.26 -> 0.28 and the whole
docling stack; torch is only probed for cuda.empty_cache()
tokenizers 0.22.2 - transformers pins <=0.23.0 and no stable 0.23.0 exists
transformers - capped <5.9.0: 5.9+ breaks docling's PDF layout model on
Apple Silicon (MPS float64), matching docling-core's own
darwin pin
docling 2.84.0 - 2.118.1 needs rapidocr >=3.9.1 and conflicts with
transformers on macOS; wants its own PR with a
golden-corpus diff
redis 7.4.0 - 8.x defaults socket_timeout to 5s, silently capping the
blocking reads in the device broker and SSE tail
websockets 16.0 - google-genai caps <17.0
marshmallow - dataclasses-json hard-caps <4; spec tightened to match
langchain block - langchain-community 0.4.2 deletes the Qdrant vectorstore
this repo imports; langchain 1.3.x needs websockets <16
Source access control
---------------------
`active_docs` is client-supplied and reached the retriever unchecked, and the
retriever queries `WHERE source_id = <id>` with no owner predicate — so any
caller could pass any source id to /stream or /api/answer and have another
tenant's documents quoted back, while /api/sources/<id>/search correctly
refused the same id. Gate it through `can_access`, the helper the guarded
endpoints already use, and filter `self.source` down to the authorized set.
Fails closed: no principal, or a check that errors, drops the source.
Three sibling paths had the same gap:
- workflow agent nodes: `AgentNodeConfig.sources` is written verbatim from
client JSON at save time and nothing validated it, so a node could name any
tenant's source. Gate against the workflow owner, so shared workflows keep
reading their owner's sources like shared agents do.
- /api/share: `_resolve_source_pg_id` resolved any id with no ownership
predicate and baked it into the agent the share creates; /api/search then
searched it. Authorize before attaching.
- search_service: re-resolve the ids stored on an agent row instead of
trusting them, so a row written by any future path with the same gap cannot
be read back.
Team grantees previously lost their source's retrieval config: the post-check
read was still owner-scoped, so it missed and fell back to defaults (an
`agentic_tool` source was bulk-prefetched for every grantee). Read unscoped
after `can_access` passes.
Retrieval
---------
`PGVectorStore._ensure_table_exists` created an IVFFlat index on the empty
table it had just created. IVFFlat computes centroids at build time, so those
centroids were random, and combined with the `source_id` post-filter a source
with hundreds of embedded chunks returned zero rows — retrieval reported no
documents, the model answered from memory, and nothing was logged. Stop
creating the index (exact search is correct and fast well past the sizes most
deployments reach); raise `ivfflat.probes` to sqrt(lists) where an index still
exists; and re-run a short indexed search exactly, since post-filtering means
no index setting can guarantee a full result. `graphrag` had the same
empty-table index with no fallback at all.
Also: bound `chunks` to 0-500 on both the request and agent paths (0 still
means "skip retrieval"), let a source's configured `retrieval.chunks` outrank
the request body, and cap ClassicRAG's per-source floor at
max(top_k, n_sources) so attaching sources cannot inflate the result set.
Silent failures
---------------
An empty retrieval was invisible to both the model and the client: the `source`
event was suppressed when the list was empty, so "searched and found nothing"
looked identical to "no source attached", and the prompt said nothing at all.
Emit the event always, and tell the model when a search ran and returned
nothing. A file that parses to nothing now fails ingest with a message naming
the cause instead of storing an embedding of the empty string. `score_threshold`
returns warnings when the active store or retriever cannot honour it.
Prompt structure
----------------
Retrieved documents move from the system prompt into the user turn, with the
injection guard restated next to them: they change every turn (defeating prefix
caching), they are third-party text that should not carry system authority, and
routing them through the query budget makes them truncatable rather than
silently crowding it out. Documents are shed lowest-ranked-first before the
question is touched.
The six chat presets (3 tones x 2 retrieval modes) differed only in their
Answering section; they are now composed from single-source fragments at load
time, not through Jinja inheritance, which would have opened a file-read
surface in the template sandbox and broken the tool-prefetch parser. Per-tool
guidance moves out of the prompt into tool schemas, so it travels with the tool
and cannot render when the tool is absent. A plain-text custom prompt is staged
as a persona value inside the skeleton instead of replacing it wholesale — it
used to silently lose the injection guard, platform block, memory and
attachments, and its braces are now inert.
Other fixes
-----------
- agents/base: an oversized system prompt drove the query budget negative and
dispatched a full-price request with an empty question; raise instead.
- llm/anthropic: migrate off the retired Text Completions API. It flattened
history to first+last message and ignored tools entirely. Adds the missing
Anthropic handler, without which every tool call was silently dropped.
- sources/upload: `sitemap` had no branch, so every sitemap ingest died on a
TypeError; `validate_url` now rejects a falsy URL cleanly.
- workflow nodes: retrieved documents never reached the node agent, so a
classic node with a source and an ordinary prompt answered "I have no
documents" while the run reported completed.
- parser/bulk: copy the metadata dict, or every chunk reports the last chunk's
token_count.
- crawler_loader: carry the page title, or citations render the whole chunk
body as the label.
Migration 0023_wiki_pages (source-scoped; version/content_hash/embed_status; FK source_id->sources ON DELETE CASCADE; UNIQUE(source_id,path) + prefix index). WikiPagesRepository: path-keyed CRUD with content-hash short-circuit, version bump, and move-with-reject. BaseVectorStore.delete_chunks_by_source_path (loop default) + parameterized pgvector override (DELETE WHERE source_id=%s AND metadata->>'source'=%s). Unit 1 of F-Wiki.
Introduces a per-source config contract that makes RAG behavior strategy-dispatched instead of a single hardcoded path. Every source gains a validated JSONB config; an empty/absent config reproduces current behavior byte-for-byte, and the whole path is gated by PER_SOURCE_RETRIEVAL_ENABLED.
Foundation: sources.config JSONB column + migration 0022_source_config; SourceConfig/ChunkingConfig/RetrievalConfig pydantic models (strict on write, lenient on read); ChunkerCreator and RetrieverCreator.register registries; config threaded through the upload routes, ingest/remote/connector workers, and reingest.
Retrieval: a Dispatcher groups sources by retriever key (all-classic collapses to today's single ClassicRAG under one shared token budget; non-classic retrievers get their own instance), removing the previous single-global-retriever collapse in stream_processor. Per-source chunks, score_threshold (honored for pgvector/mongodb, safely ignored elsewhere), and rephrase_query toggle. New PATCH /api/sources/<id>/config with team-aware (effective_write_owner) authz and a requires_reingest signal.
Chunking strategies: recursive, markdown, parent_child (selectable per source; re-ingest to apply). Search exposure: per-source prefetch vs agentic_tool for agentic/research agents. Map-reduce prescreen: optional LLM relevance pre-filter implemented as a composable post-retrieval stage that wraps any retriever.
Backend and frontend (shared Retrieval options panel + edit modal) with tests; backend suite and frontend vitest green. Excludes the wiki and GraphRAG flagships.
arc53-machine