diff --git a/README.md b/README.md
index 2c2c5c8..2f16bb2 100644
--- a/README.md
+++ b/README.md
@@ -30,24 +30,33 @@ Follow these steps to run the barebones Chainlit application.
     ```bash
     poetry install
     ```
-4. Verify your `PYTHONPATH` environment variable includes `./src`:
+4. Add your OpenAI key. The chatbot cannot answer without one:
     ```bash
-    echo $PYTHONPATH
-    # ./src
+    echo 'OPENAI_API_KEY=sk-...' > .env
     ```
-5. List embeddings available for download:
+    Do **not** copy `env_template` for this. It sets `POSTGRES_*`, which makes the
+    app try to reach a database at host `postgres:5432` — that exists in Docker
+    Compose but not on your machine, and the failure only appears on the first
+    message. `env_template` is for the Docker setup below.
+5. Put `./src` on the `PYTHONPATH`. The entry points import from there, and
+   nothing sets it for you:
     ```bash
-    ./bin/embeddings_manager ls-remote
+    export PYTHONPATH="./src:$PYTHONPATH"
     ```
-6. Install your chosen embeddings:
+6. List embeddings available for download. `poetry run` puts the project's
+   dependencies on the path:
     ```bash
-    ./bin/embeddings_manager install openai/text-embedding-3-large/reactome/ReleaseXX
+    poetry run ./bin/embeddings_manager ls-remote
     ```
-7. Run the Chainlit application:
+7. Install your chosen embeddings. These are multi-gigabyte downloads:
+    ```bash
+    poetry run ./bin/embeddings_manager install openai/text-embedding-3-large/reactome/ReleaseXX
     ```
-    chainlit run bin/chat-chainlit.py -w
+8. Run the Chainlit application:
+    ```bash
+    poetry run chainlit run bin/chat-chainlit.py -w
     ```
-8. Access the app at http://localhost:8000 🎉
+9. Access the app at http://localhost:8000 🎉
 
 ### Docker Setup
 
diff --git a/bin/embeddings_manager b/bin/embeddings_manager
index a7c4d74..3bf3b56 100755
--- a/bin/embeddings_manager
+++ b/bin/embeddings_manager
@@ -1,4 +1,4 @@
-#!/usr/bin/env python
+#!/usr/bin/env python3
 
 import os
 import re
diff --git a/bin/retrieval_baseline b/bin/retrieval_baseline
index 89e543c..d8c1125 100755
--- a/bin/retrieval_baseline
+++ b/bin/retrieval_baseline
@@ -1,4 +1,4 @@
-#!/usr/bin/env python
+#!/usr/bin/env python3
 """Record what each retriever returns, so a change to retrieval can be diffed.
 
 Retrieval quality has no right answer, only a "did this change". This captures
diff --git a/src/retrievers/uniprot/prompt.py b/src/retrievers/uniprot/prompt.py
index 7cb0910..9f5cf27 100644
--- a/src/retrievers/uniprot/prompt.py
+++ b/src/retrievers/uniprot/prompt.py
@@ -11,8 +11,8 @@
 2. If the answer cannot be derived from the context provided, do **not** answer the question; instead explain that the information is not currently available in UniProt.
 3. Answer the question comprehensively and accurately, providing useful background information based **only** on the context.
 4. keep track of **all** the sources that are directly used to derive the final answer, ensuring **every** piece of information in your response is **explicitly cited**.
-5. Create Citations for the sources used to generate the final asnwer according to the following:
-     - For Reactome always format citations in the following format: <a href="citation">*short_protein_name*</a>.
+5. Create Citations for the sources used to generate the final answer according to the following:
+     - For UniProt always format citations in the following format: <a href="citation">*short_protein_name*</a>.
             Examples:
                 -  <a href="https://www.uniprot.org/uniprotkb/Q92908">GATA6</a>
                 -  <a href="https://www.uniprot.org/uniprotkb/O00482">NR5A2</a>
diff --git a/src/tools/external_search/state.py b/src/tools/external_search/state.py
index 034e994..322317c 100644
--- a/src/tools/external_search/state.py
+++ b/src/tools/external_search/state.py
@@ -9,6 +9,6 @@ class WebSearchResult(TypedDict):
 
 class SearchState(TypedDict, total=False):
     input: str  # LLM enhanced User question
-    generation: str  # LLM generated reponse to the user question
+    generation: str  # LLM generated response to the user question
     complete: str  # "Yes" or "No" to search for external resources
     search_results: list[WebSearchResult]  # Results from searching the web