Update README.md

Author: eduard93

README.md

@@ -13,20 +13,7 @@ Examples of [PythonGateway](https://github.com/intersystems-community/PythonGate
 2. Load ObjectScript code (i.e. `do $system.OBJ.ImportDir("C:\InterSystems\Repos\PythongatewaySamples\","*.cls","c",,1)`) into Production (Ensemble-enabled) namespace. In case you want to Production-enable namespace call: `write ##class(%EnsembleMgr).EnableNamespace($Namespace, 1)`.
 3. Load restaurant data with: `do ##class(ml.match.Restaurant).Import()`.
 4. Load match data with: `do $system.OBJ.ImportDir("C:\InterSystems\Repos\PythongatewaySamples\","*.xml","c")`
-5. Install these libraries:
-
-```
-pip install recordlinkage
-pip install dedupe
-pip install pandas
-pip install numpy
-pip install matplotlib
-pip install sklearn
-pip install seaborn
-pip install tensorflow
-pip install tensorflow_hub
-pip install annoy
-```
+5. Install required libraries: `pip install -r requirements.txt`
 6. Open and start `ml.Production`.
 
 # Samples
@@ -69,19 +56,53 @@ Start `Predict Service` and `Check Service` to see model be automatically retrai
 
 ## Sample 3. Image Search engine
 
-Do not forget to switch into `Cam` category.
+This example shows image similarity search, by default we are using footwear dataset, but any type of image can be used, as long as the target object is the main object on the images in the dataset. Additionally this is how AI model can be exposed via REST API in InterSystems IRIS. This example can be used as a featuire in retail mobile app to allow user take product photos from a camera and find corresponding product(s) from the retailer shop.
+
+### Walkthrough
+
+- `ml.cam.data.Photo` class contains our dataset of photos. Use `LoadDir` method to add more photos.
+- `ml.cam.ens.bp.LoaderProcess` receives ids of `ml.cam.data.Photo` and indexes them:
+  - Images are resized and rescaled
+  - Images are passed to the chosen NN to get feature vectors (note that we use NN without the last classification layer to get better results for image as a whole)
+  - Feature Vectors are indexed using cosive index
+  - Index is saved (mmaped) on disk for later reuse
+  - Corresponding variables are saved as a `isc.py.data.Context` to Intersystems IRIS
+- On subsequent runs `ml.cam.ens.bp.LoaderProcess` first checks if saved context exists and if it does loads the index/variables instead of building it again
+- `ml.cam.ens.bs.InitService` sends initial trainig request on production startup
+- `ml.cam.ens.bp.MatcherProcess` receives one target image and uses index to find nearest matches
+- `cam.rest.Main` exposes `ml.cam.ens.bp.MatcherProcess` as a REST API webapplication `/cam`
+- `recommend.html` calls `/cam` to gest best matches for uploaded image
+
+### Docker setup
+
+To preserve the index mount volume to the `/usr/irissys/mgr/Temp` directory. `docker-compose.yml` contains an example.
+
+Test page is available at `http://localhost:52773/csp/user/recommend.html`
+
 
 ### Host setup
 
 1. Download and unpack any image dataset, for example [this](http://vision.cs.utexas.edu/projects/finegrained/utzap50k/ut-zap50k-images.zip).
 2. Load dataset into InterSystems IRIS: `zw ##class(ml.cam.data.Photo).LoadDir(<dir>)`.
 3. Restart production, it will resend initial trainig request.
-4. Training will take between 10 minutes to 5 hours depending on your CPU/GPU.
+4. Training will take 1-3 hours depending on your CPU/GPU.
 5. Trainig process saves search index automatically so subsequent runs will take <1 minute.
 6. Copy `recommend.html` to any web application.
 7. Create new unauthenticated REST app named `/cam` with `ml.cam.rest.Main` dispatch class.
 8. Open `recommend.html` in browser and send your test image
 
-### Docker setup
 
-To preserve index mount volume to the `/usr/irissys/mgr/Temp` directory. `docker-compose.yml` contains an example.
+
+### Notes
+
+- To train a new index execute:
+
+```SQL
+TRUNCATE TABLE isc_py_data.Context
+TRUNCATE TABLE isc_py_data.Variable
+```
+
+- Sometimes after trainig the index might perform poorly. In that case restart the production.
+- Prebuilt index is available [here](), corresponding globals [here](). Save index as `/usr/irissys/mgr/Temp/index.ann`  (this is Default for Docker, in general use check `WorkDirectory` value for `ml.cam.ens.bp.LoaderProcess` and save the file there as `index.ann`.
+
+