discourse-ai/spec/lib/completions/structured_output_spec.rb
Sam ab5edae121
FIX: make AI helper more robust (#1484)
* FIX: make AI helper more robust

- If JSON is broken for structured output then lean on a more forgiving parser
- Gemini 2.5 flash does not support temp, support opting out
- Evals for assistant were broken, fix interface
- Add some missing LLMs
- Translator was not mapped correctly to the feature - fix that
- Don't mix XML in prompt for translator

* lint

* correct logic

* simplify code

* implement best effort json parsing direct in the structured output object
2025-07-04 14:47:11 +10:00

158 lines
4.5 KiB
Ruby

# frozen_string_literal: true
RSpec.describe DiscourseAi::Completions::StructuredOutput do
subject(:structured_output) do
described_class.new(
{
message: {
type: "string",
},
bool: {
type: "boolean",
},
number: {
type: "integer",
},
status: {
type: "string",
},
list: {
type: "array",
items: {
type: "string",
},
},
},
)
end
describe "Parsing structured output on the fly" do
it "acts as a buffer for an streamed JSON" do
chunks = [
+"{\"message\": \"Line 1\\n",
+"Line 2\\n",
+"Line 3\", ",
+"\"bool\": true,",
+"\"number\": 4",
+"2,",
+"\"status\": \"o",
+"\\\"k\\\"\"}",
]
structured_output << chunks[0]
expect(structured_output.read_buffered_property(:message)).to eq("Line 1\n")
structured_output << chunks[1]
expect(structured_output.read_buffered_property(:message)).to eq("Line 2\n")
structured_output << chunks[2]
expect(structured_output.read_buffered_property(:message)).to eq("Line 3")
structured_output << chunks[3]
expect(structured_output.read_buffered_property(:bool)).to eq(true)
# Waiting for number to be fully buffered.
structured_output << chunks[4]
expect(structured_output.read_buffered_property(:bool)).to eq(true)
expect(structured_output.read_buffered_property(:number)).to be_nil
structured_output << chunks[5]
expect(structured_output.read_buffered_property(:number)).to eq(42)
structured_output << chunks[6]
expect(structured_output.read_buffered_property(:number)).to eq(42)
expect(structured_output.read_buffered_property(:bool)).to eq(true)
expect(structured_output.read_buffered_property(:status)).to eq("o")
structured_output << chunks[7]
expect(structured_output.read_buffered_property(:status)).to eq("\"k\"")
# No partial string left to read.
expect(structured_output.read_buffered_property(:status)).to eq("")
end
it "supports array types" do
chunks = [
+"{ \"",
+"list",
+"\":",
+" [\"",
+"Hello!",
+" I am",
+" a ",
+"chunk\",",
+"\"There\"",
+"]}",
]
structured_output << chunks[0]
structured_output << chunks[1]
structured_output << chunks[2]
expect(structured_output.read_buffered_property(:list)).to eq(nil)
structured_output << chunks[3]
expect(structured_output.read_buffered_property(:list)).to eq([""])
structured_output << chunks[4]
expect(structured_output.read_buffered_property(:list)).to eq(["Hello!"])
structured_output << chunks[5]
structured_output << chunks[6]
structured_output << chunks[7]
expect(structured_output.read_buffered_property(:list)).to eq(["Hello! I am a chunk"])
structured_output << chunks[8]
expect(structured_output.read_buffered_property(:list)).to eq(
["Hello! I am a chunk", "There"],
)
structured_output << chunks[9]
expect(structured_output.read_buffered_property(:list)).to eq(
["Hello! I am a chunk", "There"],
)
end
it "handles empty newline chunks" do
chunks = [+"{\"", +"message", +"\":\"", +"Hello!", +"\n", +"\"", +"}"]
chunks.each { |c| structured_output << c }
expect(structured_output.read_buffered_property(:message)).to eq("Hello!\n")
end
end
describe "dealing with non-JSON responses" do
it "treat it as plain text once we determined it's invalid JSON" do
chunks = [+"I'm not", +"a", +"JSON :)"]
structured_output << chunks[0]
expect(structured_output.read_buffered_property(:bob)).to eq(nil)
structured_output << chunks[1]
expect(structured_output.read_buffered_property(:bob)).to eq(nil)
structured_output << chunks[2]
structured_output.finish
expect(structured_output.read_buffered_property(:bob)).to eq(nil)
end
it "can handle broken JSON" do
broken_json = <<~JSON
```json
{
"message": "This is a broken JSON",
bool: true
}
JSON
structured_output << broken_json
structured_output.finish
expect(structured_output.read_buffered_property(:message)).to eq("This is a broken JSON")
expect(structured_output.read_buffered_property(:bool)).to eq(true)
end
end
end