diff --git a/Dockerfile b/Dockerfile
index 51b4d558c..2178f2bbd 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -1,6 +1,6 @@
 # syntax=docker/dockerfile:1.3-labs
 FROM postgres:16
-
+ENV WHERE_AM_I=docker
 ENV DEBIAN_FRONTEND=noninteractive
 USER root
 
diff --git a/test.sh b/test.sh
index fdee1a17d..f46f62040 100755
--- a/test.sh
+++ b/test.sh
@@ -8,35 +8,12 @@ if [ -f .env ]; then
   set +a
 fi
 
-if [ -z "$ENABLE_OPENAI_TESTS" ]; then
-  export ENABLE_OPENAI_TESTS=0
+if [ -n "$WHERE_AM_I" ] && [ "$WHERE_AM_I" == "docker" ]; then
+  if [ "$(whoami)" == "root" ]; then
+    echo switching to postgres user...
+    su postgres -
+  fi
+  psql -d postgres -f test.sql
+else
+  psql --no-psqlrc -d 'postgres://postgres@127.0.0.1:9876/postgres' -f test.sql
 fi
-
-if [ "$ENABLE_OPENAI_TESTS" ] && [ -z "$OPENAI_API_KEY" ]; then
-  echo "OPENAI_API_KEY must be set if running OpenAI tests"
-  exit 3
-fi
-
-if [ -z "$ENABLE_OLLAMA_TESTS" ]; then
-  export ENABLE_OLLAMA_TESTS=0
-fi
-
-if [ -z "$ENABLE_ANTHROPIC_TESTS" ]; then
-  export ENABLE_ANTHROPIC_TESTS=0
-fi
-
-if [ "$ENABLE_ANTHROPIC_TESTS" ] && [ -z "$ANTHROPIC_API_KEY" ]; then
-  echo "ANTHROPIC_API_KEY must be set if running Anthropic tests"
-  exit 3
-fi
-
-if [ -z "$ENABLE_COHERE_TESTS" ]; then
-  export ENABLE_COHERE_TESTS=0
-fi
-
-if [ "$ENABLE_COHERE_TESTS" ] && [ -z "$COHERE_API_KEY" ]; then
-  echo "COHERE_API_KEY must be set if running Cohere tests"
-  exit 3
-fi
-
-psql -d postgres -f test.sql
\ No newline at end of file
diff --git a/test.sql b/test.sql
index 8b0454706..df89b28f8 100644
--- a/test.sql
+++ b/test.sql
@@ -83,7 +83,7 @@ create table tests
 );
 
 -------------------------------------------------------------------------------
--- convenience functions for recording test results
+-- convenience function for recording test results
 create function result(_test text, _expected text, _actual text) returns bool
 as $func$
 merge into tests as t
@@ -95,57 +95,73 @@ when not matched then insert (test, actual) values (x.test, x.actual)
 select passed from tests where test = _test;
 $func$ language sql;
 
-create function result(_test text, _expected int, _actual int) returns bool
-return (select result(_test, _expected::text, _actual::text))
-;
-
-create function result(_test text, _expected bool, _actual bool) returns bool
-return (select result(_test, _expected::text, _actual::text))
-;
-
 \pset tuples_only on
 -------------------------------------------------------------------------------
 -- openai tests
 \getenv enable_openai_tests ENABLE_OPENAI_TESTS
+\if :{?enable_openai_tests}
+\else
+\set enable_openai_tests 0
+\endif
 \if :enable_openai_tests
 \set ON_ERROR_ROLLBACK on
 \set ON_ERROR_STOP off
 \i tests/openai.sql
 \set ON_ERROR_ROLLBACK off
 \set ON_ERROR_STOP on
+\else
+\echo Skipped OpenAI tests
 \endif
 
 -------------------------------------------------------------------------------
 -- ollama tests
 \getenv enable_ollama_tests ENABLE_OLLAMA_TESTS
+\if :{?enable_ollama_tests}
+\else
+\set enable_ollama_tests 0
+\endif
 \if :enable_ollama_tests
 \set ON_ERROR_ROLLBACK on
 \set ON_ERROR_STOP off
 \i tests/ollama.sql
 \set ON_ERROR_ROLLBACK off
 \set ON_ERROR_STOP on
+\else
+\echo Skipped Ollama tests
 \endif
 
 -------------------------------------------------------------------------------
 -- anthropic tests
 \getenv enable_anthropic_tests ENABLE_ANTHROPIC_TESTS
+\if :{?enable_anthropic_tests}
+\else
+\set enable_anthropic_tests 0
+\endif
 \if :enable_anthropic_tests
 \set ON_ERROR_ROLLBACK on
 \set ON_ERROR_STOP off
 \i tests/anthropic.sql
 \set ON_ERROR_ROLLBACK off
 \set ON_ERROR_STOP on
+\else
+\echo Skipped Anthropic tests
 \endif
 
 -------------------------------------------------------------------------------
 -- cohere tests
 \getenv enable_cohere_tests ENABLE_COHERE_TESTS
+\if :{?enable_cohere_tests}
+\else
+\set enable_cohere_tests 0
+\endif
 \if :enable_cohere_tests
 \set ON_ERROR_ROLLBACK on
 \set ON_ERROR_STOP off
 \i tests/cohere.sql
 \set ON_ERROR_ROLLBACK off
 \set ON_ERROR_STOP on
+\else
+\echo Skipped Cohere tests
 \endif
 
 \pset tuples_only off
@@ -153,34 +169,47 @@ return (select result(_test, _expected::text, _actual::text))
 -- test results
 \echo
 \echo
-\echo test results
+\echo
 \echo
 \echo
 
 \set ON_ERROR_STOP on
 \set ON_ERROR_ROLLBACK off
-\echo test results
-select test, passed
-from tests
-;
 
-\echo failed tests
-select *
+-- we should fail if no tests were run
+select count(*) > 0 as result
 from tests
-where passed is distinct from true
-;
+\gset
 
-\echo test stats
-select
-  count(*) as total
-, count(*) filter (where passed = true) as passed
-, count(*) filter (where passed is distinct from true) as failed
-from tests
-;
+\if :result
+
+    \echo test results
+    select test, passed
+    from tests
+    ;
+
+    \echo failed tests
+    select *
+    from tests
+    where passed is distinct from true
+    ;
+
+    \echo test stats
+    select
+      count(*) as total
+    , count(*) filter (where passed = true) as passed
+    , count(*) filter (where passed is distinct from true) as failed
+    from tests
+    ;
+
+    select count(*) filter (where passed is distinct from true) = 0 as result
+    from tests
+    \gset
+
+\else
+\warn NO TESTS WERE RUN!
+\endif
 
-select count(*) filter (where passed is distinct from true) = 0 as result
-from tests
-\gset
 
 reset role; -- no longer tester
 
diff --git a/tests/anthropic.sql b/tests/anthropic.sql
index 5259cf195..d61acd99e 100644
--- a/tests/anthropic.sql
+++ b/tests/anthropic.sql
@@ -2,6 +2,17 @@
 -- get our anthropic api key
 
 \getenv anthropic_api_key ANTHROPIC_API_KEY
+\if :{?anthropic_api_key}
+\else
+\warn Anthropic tests are enabled but ANTHROPIC_API_KEY is not set!
+do $$
+begin
+raise exception 'Anthropic tests are enabled but ANTHROPIC_API_KEY is not set!';
+end;
+$$;
+\q
+\endif
+
 -- set our session local GUC
 select set_config('ai.anthropic_api_key', $1, false) is not null as set_anthropic_api_key
 \bind :anthropic_api_key
@@ -18,7 +29,9 @@ values
 
 -------------------------------------------------------------------------------
 -- anthropic_generate
-\echo anthropic_generate
+\set testname anthropic_generate
+\set expected t
+\echo :testname
 select anthropic_generate
 ( 'claude-3-5-sonnet-20240620'
 , jsonb_build_array
@@ -32,15 +45,18 @@ select anthropic_generate
 \bind :anthropic_api_key
 \gset
 
+\if :{?actual}
 select jsonb_extract_path_text(:'actual'::jsonb, 'content', '0', 'text') is not null and (:'actual'::jsonb)->>'stop_reason' = 'end_turn' as actual
 \gset
+\endif
 
-select result('anthropic_generate', true, :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- anthropic_generate-no-key
-\echo anthropic_generate-no-key
+\set testname anthropic_generate-no-key
+\set expected t
+\echo :testname
 select anthropic_generate
 ( 'claude-3-5-sonnet-20240620'
 , jsonb_build_array
@@ -52,8 +68,9 @@ select anthropic_generate
 ) as actual
 \gset
 
+\if :{?actual}
 select jsonb_extract_path_text(:'actual'::jsonb, 'content', '0', 'text') is not null and (:'actual'::jsonb)->>'stop_reason' = 'end_turn' as actual
 \gset
+\endif
 
-select result('anthropic_generate-no-key', true, :'actual');
-\unset actual
+\ir eval.sql
diff --git a/tests/cohere.sql b/tests/cohere.sql
index 0304346c0..fcba6db2d 100644
--- a/tests/cohere.sql
+++ b/tests/cohere.sql
@@ -1,6 +1,17 @@
 -------------------------------------------------------------------------------
 -- get our cohere api key
 \getenv cohere_api_key COHERE_API_KEY
+\if :{?cohere_api_key}
+\else
+\warn Cohere tests are enabled but COHERE_API_KEY is not set!
+do $$
+begin
+raise exception 'Cohere tests are enabled but COHERE_API_KEY is not set!';
+end;
+$$;
+\q
+\endif
+
 -- set our session local GUC
 select set_config('ai.cohere_api_key', $1, false) is not null as set_cohere_api_key
 \bind :cohere_api_key
@@ -30,64 +41,81 @@ values
 
 -------------------------------------------------------------------------------
 -- cohere_list_models
-\echo cohere_list_models
-select count(*) as actual
+\set testname cohere_list_models
+\set expected t
+\echo :testname
+
+select count(*) > 0 as actual
 from cohere_list_models(_api_key=>$1)
 \bind :cohere_api_key
 \gset
 
-select result('cohere_list_models', true, :actual > 0);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_list_models-no-key
-\echo cohere_list_models-no-key
-select count(*) as actual
+\set testname cohere_list_models-no-key
+\set expected t
+\echo :testname
+
+select count(*) > 0 as actual
 from cohere_list_models()
 \gset
 
-select result('cohere_list_models-no-key', true, :actual > 0);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_list_models-endpoint
-\echo cohere_list_models-endpoint
-select count(*) as actual
+\set testname cohere_list_models-endpoint
+\set expected t
+\echo :testname
+
+select count(*) > 0 as actual
 from cohere_list_models(_endpoint=>'embed')
 \gset
 
-select result('cohere_list_models-endpoint', true, :actual > 0);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_tokenize
-\echo cohere_tokenize
-select cohere_tokenize
-( 'command'
-, 'What one programmer can do in one month, two programmers can do in two months.'
-, _api_key=>$1
+\set testname cohere_tokenize
+\set expected 17
+\echo :testname
+
+select array_length
+(
+    cohere_tokenize
+    ( 'command'
+    , 'What one programmer can do in one month, two programmers can do in two months.'
+    , _api_key=>$1
+    )
+, 1
 ) as actual
 \bind :cohere_api_key
 \gset
 
-select result('cohere_tokenize', 17, array_length(:'actual'::int[], 1));
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_tokenize-no-key
-\echo cohere_tokenize-no-key
+\set testname cohere_tokenize-no-key
+\set expected {5256,1707,1682,2383,9461,4696,1739,1863,1871,1740,9397,2112,1705,4066,3465,1742,38700,21}
+\echo :testname
+
 select cohere_tokenize
 ( 'command'
 , 'One of the best programming skills you can have is knowing when to walk away for awhile.'
 ) as actual
 \gset
 
-select result('cohere_tokenize-no-key', '{5256,1707,1682,2383,9461,4696,1739,1863,1871,1740,9397,2112,1705,4066,3465,1742,38700,21}', :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_detokenize
-\echo cohere_detokenize
+\set testname cohere_detokenize
+select 'What one programmer can do in one month, two programmers can do in two months.' as expected \gset
+\echo :testname
+
 select cohere_detokenize
 ( 'command'
 , array[5171,2011,36613,1863,1978,1703,2011,2812,19,2253,38374,1863,1978,1703,2253,3784,21]
@@ -96,34 +124,40 @@ select cohere_detokenize
 \bind :cohere_api_key
 \gset
 
-select result('cohere_detokenize', 'What one programmer can do in one month, two programmers can do in two months.', :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_detokenize-no-key
-\echo cohere_detokenize-no-key
+\set testname cohere_detokenize-no-key
+select $$Good programmers don't just write programs. They build a working vocabulary.$$ as expected \gset
+\echo :testname
+
 select cohere_detokenize
 ( 'command'
 , array[14485,38374,2630,2060,2252,5164,4905,21,2744,2628,1675,3094,23407,21]
 ) as actual
 \gset
 
-select result('cohere_detokenize-no-key', $$Good programmers don't just write programs. They build a working vocabulary.$$, :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_list_models-default-only
-\echo cohere_list_models-default-only
-select count(*) as actual
+\set testname cohere_list_models-default-only
+\set expected t
+\echo :testname
+
+select count(*) > 0 as actual
 from cohere_list_models(_endpoint=>'generate', _default_only=>true)
 \gset
 
-select result('cohere_list_models-default-only', true, :actual > 0);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_embed
-\echo cohere_embed
+\set testname cohere_embed
+\set expected 384
+\echo :testname
+
 select vector_dims
 (
     cohere_embed
@@ -136,12 +170,14 @@ select vector_dims
 \bind :cohere_api_key
 \gset
 
-select result('cohere_embed', 384, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_embed-no-key
-\echo cohere_embed-no-key
+\set testname cohere_embed-no-key
+\set expected 384
+\echo :testname
+
 select vector_dims
 (
     cohere_embed
@@ -153,12 +189,14 @@ select vector_dims
 ) as actual
 \gset
 
-select result('cohere_embed-no-key', 384, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_classify
-\echo cohere_classify
+\set testname cohere_classify
+select '{"bird": "animal", "corn": "food", "airplane": "machine"}'::jsonb::text as expected \gset
+\echo :testname
+
 with examples(example, label) as
 (
     values
@@ -180,12 +218,14 @@ from jsonb_to_recordset
 )) x(input text, prediction text)
 \gset
 
-select result('cohere_classify', '{"bird": "animal", "corn": "food", "airplane": "machine"}', :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_classify_simple
-\echo cohere_classify_simple
+\set testname cohere_classify_simple
+select '{"bird": "animal", "corn": "food", "airplane": "machine"}'::jsonb::text as expected \gset
+\echo :testname
+
 with examples(example, label) as
 (
     values
@@ -204,12 +244,13 @@ from cohere_classify_simple
 ) x
 \gset
 
-select result('cohere_classify_simple', '{"bird": "animal", "corn": "food", "airplane": "machine"}', :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_rerank
-\echo cohere_rerank
+\set testname cohere_rerank
+\set expected 2
+\echo :testname
 
 select cohere_rerank
 ( 'rerank-english-v3.0'
@@ -224,18 +265,21 @@ select cohere_rerank
 ) as actual
 \gset
 
+\if :{?actual}
 select x."index" as actual
 from jsonb_to_recordset((:'actual'::jsonb)->'results') x("index" int, "document" jsonb, relevance_score float8)
 order by relevance_score desc
 limit 1
 \gset
+\endif
 
-select result('cohere_rerank', 2, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_rerank_simple
-\echo cohere_rerank_simple
+\set testname cohere_rerank_simple
+\set expected 3
+\echo :testname
 
 select x."index" as actual
 from cohere_rerank_simple
@@ -252,12 +296,13 @@ order by relevance_score asc
 limit 1
 \gset
 
-select result('cohere_rerank_simple', 3, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- cohere_chat_complete
-\echo cohere_chat_complete
+\set testname cohere_chat_complete
+\set expected t
+\echo :testname
 
 select cohere_chat_complete
 ( 'command-r-plus'
@@ -266,7 +311,6 @@ select cohere_chat_complete
 )->>'text' is not null as actual
 \gset
 
-select result('cohere_chat_complete', true, :'actual'::bool);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
diff --git a/tests/eval.sql b/tests/eval.sql
new file mode 100644
index 000000000..0fc53429d
--- /dev/null
+++ b/tests/eval.sql
@@ -0,0 +1,6 @@
+\if :{?actual}
+select result(:'testname', :'expected', :'actual');
+\else
+select result(:'testname', :'expected', null);
+\endif
+\unset actual
diff --git a/tests/ollama.sql b/tests/ollama.sql
index fb1128c25..33a67bd5b 100644
--- a/tests/ollama.sql
+++ b/tests/ollama.sql
@@ -2,6 +2,17 @@
 -- get our ollama host
 -- grab our ollama host from the environment as a psql variable
 \getenv ollama_host OLLAMA_HOST
+\if :{?ollama_host}
+\else
+\warn Ollama tests are enabled but OLLAMA_HOST is not set!
+do $$
+begin
+raise exception 'Ollama tests are enabled but OLLAMA_HOST is not set!';
+end;
+$$;
+\q
+\endif
+
 -- set our session local GUC
 select set_config('ai.ollama_host', $1, false) is not null as set_ollama_host
 \bind :ollama_host
@@ -28,28 +39,35 @@ values
 
 -------------------------------------------------------------------------------
 -- ollama_list_models
-\echo ollama_list_models
-select count(*) as actual
+\set testname ollama_list_models
+\set expected t
+\echo :testname
+
+select count(*) > 0 as actual
 from ollama_list_models(_host=>$1)
 \bind :ollama_host
 \gset
 
-select result('ollama_list_models', true, :actual > 0);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- ollama_list_models-no-host
-\echo ollama_list_models-no-host
-select count(*) as actual
+\set testname ollama_list_models-no-host
+\set expected t
+\echo :testname
+
+select count(*) > 0 as actual
 from ollama_list_models()
 \gset
 
-select result('ollama_list_models-no-host', true, :actual > 0);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- ollama_embed
-\echo ollama_embed
+\set testname ollama_embed
+\set expected 4096
+\echo :testname
+
 select vector_dims
 (
     ollama_embed
@@ -61,12 +79,14 @@ select vector_dims
 \bind :ollama_host
 \gset
 
-select result('ollama_embed', 4096, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- ollama_embed-no-host
-\echo ollama_embed-no-host
+\set testname ollama_embed-no-host
+\set expected 4096
+\echo :testname
+
 select vector_dims
 (
     ollama_embed
@@ -76,12 +96,14 @@ select vector_dims
 ) as actual
 \gset
 
-select result('ollama_embed-no-host', 4096, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- ollama_generate
-\echo ollama_generate
+\set testname ollama_generate
+\set expected t
+\echo :testname
+
 select ollama_generate
 ( 'llama3'
 , 'what is the typical weather like in Alabama in June'
@@ -95,15 +117,19 @@ select ollama_generate
 \bind :ollama_host
 \gset
 
+\if :{?actual}
 select (:'actual'::jsonb)->>'response' is not null and ((:'actual'::jsonb)->>'done')::boolean as actual
 \gset
+\endif
 
-select result('ollama_generate', true, :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- ollama_generate-no-host
-\echo ollama_generate-no-host
+\set testname ollama_generate-no-host
+\set expected t
+\echo :testname
+
 select ollama_generate
 ( 'llama3'
 , 'what is the typical weather like in Alabama in June'
@@ -115,15 +141,19 @@ select ollama_generate
 ) as actual
 \gset
 
+\if :{?actual}
 select (:'actual'::jsonb)->>'response' is not null and ((:'actual'::jsonb)->>'done')::boolean as actual
 \gset
+\endif
 
-select result('ollama_generate-no-host', true, :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- ollama_generate-image
-\echo ollama_generate-image
+\set testname ollama_generate-image
+select 'an elephant with boxing gloves on, ready for a fight' as expected \gset
+\echo :testname
+
 select ollama_generate
 ( 'llava:7b'
 , 'Please describe this image.'
@@ -136,12 +166,19 @@ select ollama_generate
 )->>'response' as actual
 \gset
 
-select result('ollama_generate-image', 'an elephant with boxing gloves on, ready for a fight', substring(:'actual' from 152 for 52));
-\unset actual
+\if :{?actual}
+select substring(:'actual' from 152 for 52) as actual
+\gset
+\endif
+
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- ollama_chat_complete
-\echo ollama_chat_complete
+\set testname ollama_chat_complete
+\set expected t
+\echo :testname
+
 select ollama_chat_complete
 ( 'llama3'
 , jsonb_build_array
@@ -157,15 +194,19 @@ select ollama_chat_complete
 \bind :ollama_host
 \gset
 
+\if :{?actual}
 select (:'actual'::jsonb)->'message'->>'content' is not null and ((:'actual'::jsonb)->>'done')::boolean as actual
 \gset
+\endif
 
-select result('ollama_chat_complete', true, :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- ollama_chat_complete-no-host
-\echo ollama_chat_complete-no-host
+\set testname ollama_chat_complete-no-host
+\set expected t
+\echo :testname
+
 select ollama_chat_complete
 ( 'llama3'
 , jsonb_build_array
@@ -179,15 +220,19 @@ select ollama_chat_complete
 ) as actual
 \gset
 
+\if :{?actual}
 select (:'actual'::jsonb)->'message'->>'content' is not null and ((:'actual'::jsonb)->>'done')::boolean as actual
 \gset
+\endif
 
-select result('ollama_chat_complete-no-host', true, :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- ollama_chat_complete-image
-\echo ollama_chat_complete-image
+\set testname ollama_chat_complete-image
+\set expected t
+\echo :testname
+
 select ollama_chat_complete
 ( 'llava:7b'
 , jsonb_build_array
@@ -204,26 +249,34 @@ select ollama_chat_complete
 )->'message'->>'content' as actual
 \gset
 
-select result('ollama_chat_complete-image', true, starts_with(:'actual'::text, ' This is a digitally manipulated image'));
-\unset actual
+\if :{?actual}
+select starts_with(:'actual'::text, ' This is a digitally manipulated image') as actual
+\gset
+\endif
+
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- ollama_ps
-\echo ollama_ps
+\set testname ollama_ps
+\set expected 1
+\echo :testname
+
 select count(*) filter (where "name" = 'llava:7b') as actual
 from ollama_ps(_host=>$1)
 \bind :ollama_host
 \gset
 
-select result('ollama_ps', 1, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- ollama_ps-no-host
-\echo ollama_ps-no-host
+\set testname ollama_ps-no-host
+\set expected 1
+\echo :testname
+
 select count(*) filter (where "name" = 'llava:7b') as actual
 from ollama_ps()
 \gset
 
-select result('ollama_ps-no-host', 1, :actual);
-\unset actual
+\ir eval.sql
diff --git a/tests/openai.sql b/tests/openai.sql
index e29ae13c7..48be64440 100644
--- a/tests/openai.sql
+++ b/tests/openai.sql
@@ -2,6 +2,17 @@
 -- get our openai api key
 -- grab our api key from the environment as a psql variable
 \getenv openai_api_key OPENAI_API_KEY
+\if :{?openai_api_key}
+\else
+\warn OpenAI tests are enabled but OPENAI_API_KEY is not set!
+do $$
+begin
+raise exception 'OpenAI tests are enabled but OPENAI_API_KEY is not set!';
+end;
+$$;
+\q
+\endif
+
 -- set our session local GUC
 select set_config('ai.openai_api_key', $1, false) is not null as set_openai_api_key
 \bind :openai_api_key
@@ -34,46 +45,57 @@ values
 
 -------------------------------------------------------------------------------
 -- openai_list_models
-\echo openai_list_models
-select count(*) as actual
+\set testname openai_list_models
+\set expected t
+\echo :testname
+
+select count(*) > 0 as actual
 from openai_list_models(_api_key=>$1)
 \bind :openai_api_key
 \gset
 
-select result('openai_list_models', true, :actual > 0);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_list_models-no-key
-\echo openai_list_models-no-key
-select count(*) as actual
+\set testname openai_list_models-no-key
+\set expected t
+\echo :testname
+
+select count(*) > 0 as actual
 from openai_list_models()
 \gset
 
-select result('openai_list_models-no-key', true, :actual > 0);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_tokenize
-\echo openai_tokenize
-select openai_tokenize('text-embedding-ada-002', 'the purple elephant sits on a red mushroom') as actual
+\set testname openai_tokenize
+select array[1820,25977,46840,23874,389,264,2579,58466]::text as expected \gset
+\echo :testname
+
+select openai_tokenize('text-embedding-ada-002', 'the purple elephant sits on a red mushroom')::text as actual
 \gset
 
-select result('openai_tokenize', array[1820,25977,46840,23874,389,264,2579,58466]::text, :'actual'::int[]::text);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_detokenize
-\echo openai_detokenize
+\set testname openai_detokenize
+select 'the purple elephant sits on a red mushroom' as expected \gset
+\echo :testname
+
 select openai_detokenize('text-embedding-ada-002', array[1820,25977,46840,23874,389,264,2579,58466]) as actual
 \gset
 
-select result('openai_detokenize', 'the purple elephant sits on a red mushroom', :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_embed-1
-\echo openai_embed-1
+\set testname openai_embed-1
+\set expected 1536
+\echo :testname
+
 select vector_dims
 (
     openai_embed
@@ -85,12 +107,14 @@ select vector_dims
 \bind :openai_api_key
 \gset
 
-select result('openai_embed-1', 1536, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_embed-1-no-key
-\echo openai_embed-1-no-key
+\set testname openai_embed-1-no-key
+\set expected 1536
+\echo :testname
+
 select vector_dims
 (
     openai_embed
@@ -100,12 +124,14 @@ select vector_dims
 ) as actual
 \gset
 
-select result('openai_embed-1-no-key', 1536, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_embed-2
-\echo openai_embed-2
+\set testname openai_embed-2
+\set expected 768
+\echo :testname
+
 select vector_dims
 (
     openai_embed
@@ -118,12 +144,14 @@ select vector_dims
 \bind :openai_api_key
 \gset
 
-select result('openai_embed-2', 768, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_embed-2-no-key
-\echo openai_embed-2-no-key
+\set testname openai_embed-2-no-key
+\set expected 768
+\echo :testname
+
 select vector_dims
 (
     openai_embed
@@ -134,12 +162,14 @@ select vector_dims
 ) as actual
 \gset
 
-select result('openai_embed-2-no-key', 768, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_embed-3
-\echo openai_embed-3
+\set testname openai_embed-3
+\set expected 3072
+\echo :testname
+
 select vector_dims
 (
     openai_embed
@@ -152,12 +182,14 @@ select vector_dims
 \bind :openai_api_key
 \gset
 
-select result('openai_embed-3', 3072, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_embed-3-no-key
-\echo openai_embed-3-no-key
+\set testname openai_embed-3-no-key
+\set expected 3072
+\echo :testname
+
 select vector_dims
 (
     openai_embed
@@ -168,12 +200,14 @@ select vector_dims
 ) as actual
 \gset
 
-select result('openai_embed-3-no-key', 3072, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_embed-4
-\echo openai_embed-4
+\set testname openai_embed-4
+\set expected 6144
+\echo :testname
+
 select sum(vector_dims(embedding)) as actual
 from openai_embed
 ( 'text-embedding-3-large'
@@ -183,12 +217,14 @@ from openai_embed
 \bind :openai_api_key
 \gset
 
-select result('openai_embed-4', 6144, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_embed-4-no-key
-\echo openai_embed-4-no-key
+\set testname openai_embed-4-no-key
+\set expected 6144
+\echo :testname
+
 select sum(vector_dims(embedding)) as actual
 from openai_embed
 ( 'text-embedding-3-large'
@@ -196,12 +232,14 @@ from openai_embed
 )
 \gset
 
-select result('openai_embed-4-no-key', 6144, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_embed-5
-\echo openai_embed-5
+\set testname openai_embed-5
+\set expected 1536
+\echo :testname
+
 select vector_dims
 (
     openai_embed
@@ -213,12 +251,14 @@ select vector_dims
 \bind :openai_api_key
 \gset
 
-select result('openai_embed-5', 1536, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_embed-5-no-key
-\echo openai_embed-5-no-key
+\set testname openai_embed-5-no-key
+\set expected 1536
+\echo :testname
+
 select vector_dims
 (
     openai_embed
@@ -228,12 +268,14 @@ select vector_dims
 ) as actual
 \gset
 
-select result('openai_embed-5-no-key', 1536, :actual);
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_chat_complete
-\echo openai_chat_complete
+\set testname openai_chat_complete
+\set expected t
+\echo :testname
+
 select openai_chat_complete
 ( 'gpt-4o'
 , jsonb_build_array
@@ -245,15 +287,19 @@ select openai_chat_complete
 \bind :openai_api_key
 \gset
 
+\if :{?actual}
 select jsonb_extract_path_text(:'actual'::jsonb, 'choices', '0', 'message', 'content') is not null as actual
 \gset
+\endif
 
-select result('openai_chat_complete', true, :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_chat_complete-no-key
-\echo openai_chat_complete-no-key
+\set testname openai_chat_complete-no-key
+\set expected t
+\echo :testname
+
 select openai_chat_complete
 ( 'gpt-4o'
 , jsonb_build_array
@@ -263,15 +309,19 @@ select openai_chat_complete
 ) as actual
 \gset
 
+\if :{?actual}
 select jsonb_extract_path_text(:'actual'::jsonb, 'choices', '0', 'message', 'content') is not null as actual
 \gset
+\endif
 
-select result('openai_chat_complete-no-key', true, :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_moderate
-\echo openai_moderate
+\set testname openai_moderate
+\set expected t
+\echo :testname
+
 select openai_moderate
 ( 'text-moderation-stable'
 , 'I want to kill them.'
@@ -280,25 +330,30 @@ select openai_moderate
 \bind :openai_api_key
 \gset
 
+\if :{?actual}
 select jsonb_extract_path_text(:'actual'::jsonb, 'results', '0', 'flagged')::bool as actual
 \gset
+\endif
 
-select result('openai_moderate', true, :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------
 -- openai_moderate-no-key
-\echo openai_moderate-no-key
+\set testname openai_moderate-no-key
+\set expected t
+\echo :testname
+
 select openai_moderate
 ( 'text-moderation-stable'
 , 'I want to kill them.'
 ) as actual
 \gset
 
+\if :{?actual}
 select jsonb_extract_path_text(:'actual'::jsonb, 'results', '0', 'flagged')::bool as actual
 \gset
+\endif
 
-select result('openai_moderate-no-key', true, :'actual');
-\unset actual
+\ir eval.sql
 
 -------------------------------------------------------------------------------