The image was 2.45GB, of which the venv is 1.73GB. The build was already
multi-stage and already installed CPU-only torch, so the remaining weight was
not build tooling - it was payload inside the installed packages that the
running service can never execute.
Removed in the build stage, before the runtime stage copies /opt/venv, so the
bytes never enter the final image:
- bundled test suites (~237MB; torch/test is 83MB, pandas/tests 40MB)
- torch/include (62MB), C++ headers for compiling against libtorch
- torch/bin (50MB), gtest binaries and protoc; torch_shm_manager is kept
pytest and httpx2 move to requirements-dev.txt: the image copies app/, cli/,
scripts/, data/ and serve.py, never tests/, so the test stack was unusable
there regardless.
sympy was checked and deliberately kept - 'import sentence_transformers' does
pull it in through torch.fx, so removing it would break embeddings.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The deployment landed on mcp.nearle.ai.in, not the mcp.catalogue.nearle.ai.in
these references were written against. Only comments, README and .env.example
are affected - nothing reads the hostname at runtime - but a wrong host in the
connection snippet is a wrong host somebody pastes into an MCP client.
API_CORS_ORIGINS is unchanged: it takes the FRONTEND's origin
(catalogue.nearle.ai.in), not the API's, so moving the API does not affect it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Adds a FastMCP server mounted onto the existing FastAPI app, so it ships in the
same container and answers on the same host rather than needing a process of
its own.
Fifteen tools, all read-only: catalog search and browse, nutrition facts and
health scores, healthier alternatives, per-store inventory and pricing,
discounts, trending and sales analytics. Each wraps a service function the REST
API already reaches through a GET. None of the write or compute endpoints are
exposed, because a tool list is chosen from by a model rather than by a person,
and catalog generation or model retraining is not something to leave one tool
call away.
The tools are written by hand rather than generated from the OpenAPI schema.
Mirroring all 63 routes would work, but a model picks a tool by reading its
description, and 63 near-identical generated entries is a worse thing to choose
from than a dozen written to be told apart.
Authentication reuses the access token from POST /api/auth/login - no separate
MCP credential, the same Principal and expiry as the REST API. The check lives
in one middleware rather than at the top of each tool, so a tool added later
cannot be left unguarded by forgetting a line. Note that this requires passing
include={"authorization"} to get_http_headers(), which strips that header by
default to avoid forwarding it downstream; without it the header is invisible
and every request looks unauthenticated, valid ones included.
Mounting a sub-app does not run its lifespan - only the outermost app's is
executed - so the MCP app's lifespan is chained through the FastAPI one. Without
that the endpoint accepts a connection and then fails on the first message with
a session manager that was never started.
Adds GET /api/mcp/info and POST /api/mcp/tools/{name} for the admin UI. The MCP
endpoint speaks streamable HTTP with session handling, so rendering a tool list
in the browser would otherwise mean shipping a full MCP client in React.
fastmcp needs Python 3.10+. The image is 3.11; on anything older the import
fails and the REST API starts without the MCP endpoint instead of not starting.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Dokploy routes the domain to port 3000, but the container only bound 8000, so
the proxy had nothing to talk to and the domain returned 502 with a perfectly
healthy process behind it.
The frontend image already solved this by answering on both 80 and 3000
(`listen 80; listen 3000;` in nginx.conf). Do the same here rather than swap one
guess for another: 3000 is what the platform routes to, and 8000 is what the
README, the vite dev proxy and docker-compose all target, so binding both means
the container works whichever one it is pointed at.
uvicorn's CLI takes a single --port, but Server.run() accepts pre-bound
sockets, so serve.py binds each port and hands the list to one uvicorn - no
extra worker or second process to supervise. PORT still pins a single port for
anyone who wants one; PORTS changes the pair.
A port that cannot be bound is logged and skipped rather than being fatal,
since losing one of the two should not take down a service the platform only
routes to on the other. It exits non-zero only when nothing is listening at
all, so a genuinely dead container is still reported as failed.
The healthcheck moves into the same file and passes if either port answers,
which keeps it from drifting out of sync with what is actually bound.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
/api/upload/nutrition called json.dumps() in a module that never imported
json, so every request to it raised NameError, was swallowed by the broad
except, and came back as "500 Database import failed". Import json.
Persist the three directories the app writes to at runtime. Products added
through the UI are appended to data/seed_catalogs/*.json and retrained models
are written to app/intelligence/artifacts/*.joblib; both live inside the image,
so a redeploy silently discarded them. The paths now come from settings
(DATA_DIR / SEED_CATALOG_DIR / MODEL_ARTIFACTS_DIR) so a volume can be mounted
on them, and catalog_engine.save_catalog resolves against DATA_DIR instead of
a working-directory-relative "data/", which landed somewhere different
depending on where the process was started from.
Mounting those volumes would otherwise have made things worse: Docker seeds a
named volume from the image on first use, but a bind mount starts empty and
just hides what the image shipped. A bind mount on /app/data would have left
the API with no seed catalogs, so the next product added would write a JSON
file containing only that product. The image now keeps pristine copies at
/app/.bundled, and restore_bundled_assets() tops up whatever a freshly mounted
directory is missing at startup without overwriting anything already there.
Configure CORS for the split-domain deployment: the React app is served from
catalogue.nearle.ai.in and calls the API on mcp.catalogue.nearle.ai.in, so the
frontend origin has to be in API_CORS_ORIGINS. A wrong list fails only in the
browser while the server logs a healthy 200, so the effective origins are now
logged at startup with a warning when they are localhost-only.
Fix FRONTEND_DIST, which looked for a sibling "frontend/" directory that is
actually named "catalogue_frontend/", so the single-port unified-serving branch
could never activate even with a build sitting next to it.
Rebuild the Dockerfile on the frontend's multi-stage pattern: dependencies
resolve into a venv in a build stage, the runtime stage copies only that.
Adds PYTHONUNBUFFERED so startup errors reach Dokploy's log pane, a liveness
HEALTHCHECK (/api/health answers 200 even when Postgres is down, so a database
blip cannot restart-loop the container), and an overridable PORT. The CMD execs
uvicorn so SIGTERM reaches it rather than the sh wrapper.
Add "from __future__ import annotations" to ollama_service and image_search,
which used PEP 604 unions in runtime-evaluated signatures and so could not be
imported below Python 3.10.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
