|
1 | 1 | package evaluation |
2 | 2 |
|
3 | 3 | import ( |
| 4 | + "encoding/json" |
| 5 | + "os" |
4 | 6 | "path/filepath" |
5 | 7 | "testing" |
6 | 8 | "time" |
@@ -107,6 +109,92 @@ func TestSaveRunSessions(t *testing.T) { |
107 | 109 | assert.True(t, titles["eval-test-2"], "should have eval-test-2") |
108 | 110 | } |
109 | 111 |
|
| 112 | +func TestSaveRunSessionsJSON(t *testing.T) { |
| 113 | + t.Parallel() |
| 114 | + |
| 115 | + outputDir := t.TempDir() |
| 116 | + |
| 117 | + // Create sessions with different content |
| 118 | + sess1 := session.New( |
| 119 | + session.WithTitle("eval-json-1"), |
| 120 | + session.WithUserMessage("What is the capital of France?"), |
| 121 | + ) |
| 122 | + sess1.InputTokens = 100 |
| 123 | + sess1.OutputTokens = 50 |
| 124 | + sess1.Cost = 0.01 |
| 125 | + |
| 126 | + sess2 := session.New( |
| 127 | + session.WithTitle("eval-json-2"), |
| 128 | + session.WithUserMessage("What is 2+2?"), |
| 129 | + ) |
| 130 | + sess2.InputTokens = 80 |
| 131 | + sess2.OutputTokens = 30 |
| 132 | + sess2.Cost = 0.005 |
| 133 | + |
| 134 | + // Create an eval run with sessions |
| 135 | + run := &EvalRun{ |
| 136 | + Name: "test-json-001", |
| 137 | + Timestamp: time.Now(), |
| 138 | + Results: []Result{ |
| 139 | + { |
| 140 | + Title: "eval-json-1", |
| 141 | + Question: "What is the capital of France?", |
| 142 | + Response: "Paris is the capital of France.", |
| 143 | + Session: sess1, |
| 144 | + }, |
| 145 | + { |
| 146 | + Title: "eval-json-2", |
| 147 | + Question: "What is 2+2?", |
| 148 | + Response: "4", |
| 149 | + Session: sess2, |
| 150 | + }, |
| 151 | + { |
| 152 | + // Result without a session (error case) |
| 153 | + Title: "eval-json-3", |
| 154 | + Error: "container failed", |
| 155 | + Session: nil, |
| 156 | + }, |
| 157 | + }, |
| 158 | + } |
| 159 | + |
| 160 | + // Save sessions to JSON |
| 161 | + sessionsPath, err := SaveRunSessionsJSON(run, outputDir) |
| 162 | + require.NoError(t, err) |
| 163 | + assert.Equal(t, filepath.Join(outputDir, "test-json-001.json"), sessionsPath) |
| 164 | + assert.FileExists(t, sessionsPath) |
| 165 | + |
| 166 | + // Read and parse the JSON file |
| 167 | + data, err := os.ReadFile(sessionsPath) |
| 168 | + require.NoError(t, err) |
| 169 | + |
| 170 | + var loadedSessions []*session.Session |
| 171 | + err = json.Unmarshal(data, &loadedSessions) |
| 172 | + require.NoError(t, err) |
| 173 | + |
| 174 | + // Should have 2 sessions (excluding the error case) |
| 175 | + assert.Len(t, loadedSessions, 2) |
| 176 | + |
| 177 | + // Verify session content |
| 178 | + titles := make(map[string]*session.Session) |
| 179 | + for _, sess := range loadedSessions { |
| 180 | + titles[sess.Title] = sess |
| 181 | + } |
| 182 | + |
| 183 | + assert.Contains(t, titles, "eval-json-1") |
| 184 | + assert.Contains(t, titles, "eval-json-2") |
| 185 | + |
| 186 | + // Verify cost and token data is preserved |
| 187 | + sess1Loaded := titles["eval-json-1"] |
| 188 | + assert.Equal(t, int64(100), sess1Loaded.InputTokens) |
| 189 | + assert.Equal(t, int64(50), sess1Loaded.OutputTokens) |
| 190 | + assert.InDelta(t, 0.01, sess1Loaded.Cost, 0.0001) |
| 191 | + |
| 192 | + sess2Loaded := titles["eval-json-2"] |
| 193 | + assert.Equal(t, int64(80), sess2Loaded.InputTokens) |
| 194 | + assert.Equal(t, int64(30), sess2Loaded.OutputTokens) |
| 195 | + assert.InDelta(t, 0.005, sess2Loaded.Cost, 0.0001) |
| 196 | +} |
| 197 | + |
110 | 198 | func TestSaveRunSessionsWithCost(t *testing.T) { |
111 | 199 | t.Parallel() |
112 | 200 |
|
|
0 commit comments