Skip to content

Tru

Bases: SingletonPerName

Tru is the main class that provides an entry points to trulens-eval. Tru lets you:

  • Log chain prompts and outputs
  • Log chain Metadata
  • Run and log feedback functions
  • Run streamlit dashboard to view experiment results

All data is logged to the current working directory to default.sqlite.

Source code in trulens_eval/trulens_eval/tru.py
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
class Tru(SingletonPerName):
    """
    Tru is the main class that provides an entry points to trulens-eval. Tru lets you:

    * Log chain prompts and outputs
    * Log chain Metadata
    * Run and log feedback functions
    * Run streamlit dashboard to view experiment results

    All data is logged to the current working directory to default.sqlite.
    """
    DEFAULT_DATABASE_FILE = "default.sqlite"

    # Process or Thread of the deferred feedback function evaluator.
    evaluator_proc = None

    # Process of the dashboard app.
    dashboard_proc = None

    def Chain(self, *args, **kwargs):
        """
        Create a TruChain with database managed by self.
        """

        from trulens_eval.tru_chain import TruChain

        return TruChain(tru=self, *args, **kwargs)

    def __init__(self):
        """
        TruLens instrumentation, logging, and feedback functions for chains.
        Creates a local database 'default.sqlite' in current working directory.
        """

        if hasattr(self, "db"):
            # Already initialized by SingletonByName mechanism.
            return

        self.db = LocalSQLite(Tru.DEFAULT_DATABASE_FILE)

    def reset_database(self):
        """
        Reset the database. Clears all tables.
        """

        self.db.reset_database()

    def add_record(
        self,
        prompt: str,
        response: str,
        record_json: JSON,
        tags: Optional[str] = "",
        ts: Optional[int] = None,
        total_tokens: Optional[int] = None,
        total_cost: Optional[float] = None,
    ):
        """
        Add a record to the database.

        Parameters:

            prompt (str): Chain input or "prompt".

            response (str): Chain output or "response".

            record_json (JSON): Record as produced by `TruChain.call_with_record`.

            tags (str, optional): Additional metadata to include with the record.

            ts (int, optional): Timestamp of record creation.

            total_tokens (int, optional): The number of tokens generated in
            producing the response.

            total_cost (float, optional): The cost of producing the response.

        Returns:
            str: Unique record identifier.

        """
        ts = ts or datetime.now()
        total_tokens = total_tokens or record_json['_cost']['total_tokens']
        total_cost = total_cost or record_json['_cost']['total_cost']

        chain_id = record_json['chain_id']

        record_id = self.db.insert_record(
            chain_id=chain_id,
            input=prompt,
            output=response,
            record_json=record_json,
            ts=ts,
            tags=tags,
            total_tokens=total_tokens,
            total_cost=total_cost
        )

        return record_id

    def run_feedback_functions(
        self,
        record_json: JSON,
        feedback_functions: Sequence['Feedback'],
        chain_json: Optional[JSON] = None,
    ) -> Sequence[JSON]:
        """
        Run a collection of feedback functions and report their result.

        Parameters:

            record_json (JSON): The record on which to evaluate the feedback
            functions.

            chain_json (JSON, optional): The chain that produced the given record.
            If not provided, it is looked up from the given database `db`.

            feedback_functions (Sequence[Feedback]): A collection of feedback
            functions to evaluate.

        Returns nothing.
        """

        chain_id = record_json['chain_id']

        if chain_json is None:
            chain_json = self.db.get_chain(chain_id=chain_id)
            if chain_json is None:
                raise RuntimeError(
                    "Chain {chain_id} not present in db. "
                    "Either add it with `tru.add_chain` or provide `chain_json` to `tru.run_feedback_functions`."
                )

        else:
            assert chain_id == chain_json[
                'chain_id'], "Record was produced by a different chain."

            if self.db.get_chain(chain_id=chain_json['chain_id']) is None:
                logging.warn(
                    "Chain {chain_id} was not present in database. Adding it."
                )
                self.add_chain(chain_json=chain_json)

        evals = []

        for func in feedback_functions:
            evals.append(
                TP().promise(
                    lambda f: f.run_on_record(
                        chain_json=chain_json, record_json=record_json
                    ), func
                )
            )

        evals = map(lambda p: p.get(), evals)

        return list(evals)

    def add_chain(
        self, chain_json: JSON, chain_id: Optional[str] = None
    ) -> None:
        """
        Add a chain to the database.        
        """

        self.db.insert_chain(chain_id=chain_id, chain_json=chain_json)

    def add_feedback(self, result_json: JSON) -> None:
        """
        Add a single feedback result to the database.
        """

        if 'record_id' not in result_json or result_json['record_id'] is None:
            raise RuntimeError(
                "Result does not include record_id. "
                "To log feedback, log the record first using `tru.add_record`."
            )

        self.db.insert_feedback(result_json=result_json, status=2)

    def add_feedbacks(self, result_jsons: Iterable[JSON]) -> None:
        """
        Add multiple feedback results to the database.
        """

        for result_json in result_jsons:
            self.add_feedback(result_json=result_json)

    def get_chain(self, chain_id: str) -> JSON:
        """
        Look up a chain from the database.
        """

        return self.db.get_chain(chain_id)

    def get_records_and_feedback(self, chain_ids: List[str]):
        """
        Get records, their feeback results, and feedback names from the database.
        """

        df, feedback_columns = self.db.get_records_and_feedback(chain_ids)

        return df, feedback_columns

    def start_evaluator(self, fork=False) -> Union[Process, Thread]:
        """
        Start a deferred feedback function evaluation thread.
        """

        assert not fork, "Fork mode not yet implemented."

        if self.evaluator_proc is not None:
            raise RuntimeError("Evaluator is already running in this process.")

        from trulens_eval.tru_feedback import Feedback

        if not fork:
            self.evaluator_stop = threading.Event()

        def runloop():
            while fork or not self.evaluator_stop.is_set():
                print("Looking for things to do. Stop me with `tru.stop_evaluator()`.", end='')
                Feedback.evaluate_deferred(tru=self)
                TP().finish(timeout=10)
                if fork:
                    sleep(10)
                else:
                    self.evaluator_stop.wait(10)

            print("Evaluator stopped.")

        if fork:
            proc = Process(target=runloop)
        else:
            proc = Thread(target=runloop)

        # Start a persistent thread or process that evaluates feedback functions.

        self.evaluator_proc = proc
        proc.start()

        return proc

    def stop_evaluator(self):
        """
        Stop the deferred feedback evaluation thread.
        """

        if self.evaluator_proc is None:
            raise RuntimeError("Evaluator not running this process.")

        if isinstance(self.evaluator_proc, Process):
            self.evaluator_proc.terminate()

        elif isinstance(self.evaluator_proc, Thread):
            self.evaluator_stop.set()
            self.evaluator_proc.join()
            self.evaluator_stop = None

        self.evaluator_proc = None

    def stop_dashboard(self) -> None:
        """Stop existing dashboard if running.

        Raises:
            ValueError: Dashboard is already running.
        """
        if Tru.dashboard_proc is None:
            raise ValueError("Dashboard not running.")

        Tru.dashboard_proc.kill()
        Tru.dashboard_proc = None

    def run_dashboard(self, _dev: bool = False) -> Process:
        """ Runs a streamlit dashboard to view logged results and chains

        Raises:
            ValueError: Dashboard is already running.

        Returns:
            Process: Process containing streamlit dashboard.
        """

        if Tru.dashboard_proc is not None:
            raise ValueError("Dashboard already running. Run tru.stop_dashboard() to stop existing dashboard.")

        # Create .streamlit directory if it doesn't exist
        streamlit_dir = os.path.join(os.getcwd(), '.streamlit')
        os.makedirs(streamlit_dir, exist_ok=True)

        # Create config.toml file
        config_path = os.path.join(streamlit_dir, 'config.toml')
        with open(config_path, 'w') as f:
            f.write('[theme]\n')
            f.write('primaryColor="#0A2C37"\n')
            f.write('backgroundColor="#FFFFFF"\n')
            f.write('secondaryBackgroundColor="F5F5F5"\n')
            f.write('textColor="#0A2C37"\n')
            f.write('font="sans serif"\n')

        cred_path = os.path.join(streamlit_dir, 'credentials.toml')
        with open(cred_path, 'w') as f:
            f.write('[general]\n')
            f.write('email=""\n')

        #run leaderboard with subprocess
        leaderboard_path = pkg_resources.resource_filename(
            'trulens_eval', 'Leaderboard.py'
        )

        env_opts = {}
        if _dev:
            env_opts['env'] = os.environ
            env_opts['env']['PYTHONPATH'] = str(Path.cwd())

        proc = subprocess.Popen(
            ["streamlit", "run", "--server.headless=True", leaderboard_path], **env_opts
        )

        Tru.dashboard_proc = proc

        return proc

    start_dashboard = run_dashboard

Chain(*args, **kwargs)

Create a TruChain with database managed by self.

Source code in trulens_eval/trulens_eval/tru.py
40
41
42
43
44
45
46
47
def Chain(self, *args, **kwargs):
    """
    Create a TruChain with database managed by self.
    """

    from trulens_eval.tru_chain import TruChain

    return TruChain(tru=self, *args, **kwargs)

__init__()

TruLens instrumentation, logging, and feedback functions for chains. Creates a local database 'default.sqlite' in current working directory.

Source code in trulens_eval/trulens_eval/tru.py
49
50
51
52
53
54
55
56
57
58
59
def __init__(self):
    """
    TruLens instrumentation, logging, and feedback functions for chains.
    Creates a local database 'default.sqlite' in current working directory.
    """

    if hasattr(self, "db"):
        # Already initialized by SingletonByName mechanism.
        return

    self.db = LocalSQLite(Tru.DEFAULT_DATABASE_FILE)

add_chain(chain_json, chain_id=None)

Add a chain to the database.

Source code in trulens_eval/trulens_eval/tru.py
179
180
181
182
183
184
185
186
def add_chain(
    self, chain_json: JSON, chain_id: Optional[str] = None
) -> None:
    """
    Add a chain to the database.        
    """

    self.db.insert_chain(chain_id=chain_id, chain_json=chain_json)

add_feedback(result_json)

Add a single feedback result to the database.

Source code in trulens_eval/trulens_eval/tru.py
188
189
190
191
192
193
194
195
196
197
198
199
def add_feedback(self, result_json: JSON) -> None:
    """
    Add a single feedback result to the database.
    """

    if 'record_id' not in result_json or result_json['record_id'] is None:
        raise RuntimeError(
            "Result does not include record_id. "
            "To log feedback, log the record first using `tru.add_record`."
        )

    self.db.insert_feedback(result_json=result_json, status=2)

add_feedbacks(result_jsons)

Add multiple feedback results to the database.

Source code in trulens_eval/trulens_eval/tru.py
201
202
203
204
205
206
207
def add_feedbacks(self, result_jsons: Iterable[JSON]) -> None:
    """
    Add multiple feedback results to the database.
    """

    for result_json in result_jsons:
        self.add_feedback(result_json=result_json)

add_record(prompt, response, record_json, tags='', ts=None, total_tokens=None, total_cost=None)

Add a record to the database.

Parameters:

Name Type Description Default
prompt str

Chain input or "prompt".

required
response str

Chain output or "response".

required
record_json JSON

Record as produced by TruChain.call_with_record.

required
tags str

Additional metadata to include with the record.

''
ts int

Timestamp of record creation.

None
total_tokens int

The number of tokens generated in

None
total_cost float

The cost of producing the response.

None

Returns:

Name Type Description
str

Unique record identifier.

Source code in trulens_eval/trulens_eval/tru.py
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
def add_record(
    self,
    prompt: str,
    response: str,
    record_json: JSON,
    tags: Optional[str] = "",
    ts: Optional[int] = None,
    total_tokens: Optional[int] = None,
    total_cost: Optional[float] = None,
):
    """
    Add a record to the database.

    Parameters:

        prompt (str): Chain input or "prompt".

        response (str): Chain output or "response".

        record_json (JSON): Record as produced by `TruChain.call_with_record`.

        tags (str, optional): Additional metadata to include with the record.

        ts (int, optional): Timestamp of record creation.

        total_tokens (int, optional): The number of tokens generated in
        producing the response.

        total_cost (float, optional): The cost of producing the response.

    Returns:
        str: Unique record identifier.

    """
    ts = ts or datetime.now()
    total_tokens = total_tokens or record_json['_cost']['total_tokens']
    total_cost = total_cost or record_json['_cost']['total_cost']

    chain_id = record_json['chain_id']

    record_id = self.db.insert_record(
        chain_id=chain_id,
        input=prompt,
        output=response,
        record_json=record_json,
        ts=ts,
        tags=tags,
        total_tokens=total_tokens,
        total_cost=total_cost
    )

    return record_id

get_chain(chain_id)

Look up a chain from the database.

Source code in trulens_eval/trulens_eval/tru.py
209
210
211
212
213
214
def get_chain(self, chain_id: str) -> JSON:
    """
    Look up a chain from the database.
    """

    return self.db.get_chain(chain_id)

get_records_and_feedback(chain_ids)

Get records, their feeback results, and feedback names from the database.

Source code in trulens_eval/trulens_eval/tru.py
216
217
218
219
220
221
222
223
def get_records_and_feedback(self, chain_ids: List[str]):
    """
    Get records, their feeback results, and feedback names from the database.
    """

    df, feedback_columns = self.db.get_records_and_feedback(chain_ids)

    return df, feedback_columns

reset_database()

Reset the database. Clears all tables.

Source code in trulens_eval/trulens_eval/tru.py
61
62
63
64
65
66
def reset_database(self):
    """
    Reset the database. Clears all tables.
    """

    self.db.reset_database()

run_dashboard(_dev=False)

Runs a streamlit dashboard to view logged results and chains

Raises:

Type Description
ValueError

Dashboard is already running.

Returns:

Name Type Description
Process Process

Process containing streamlit dashboard.

Source code in trulens_eval/trulens_eval/tru.py
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
def run_dashboard(self, _dev: bool = False) -> Process:
    """ Runs a streamlit dashboard to view logged results and chains

    Raises:
        ValueError: Dashboard is already running.

    Returns:
        Process: Process containing streamlit dashboard.
    """

    if Tru.dashboard_proc is not None:
        raise ValueError("Dashboard already running. Run tru.stop_dashboard() to stop existing dashboard.")

    # Create .streamlit directory if it doesn't exist
    streamlit_dir = os.path.join(os.getcwd(), '.streamlit')
    os.makedirs(streamlit_dir, exist_ok=True)

    # Create config.toml file
    config_path = os.path.join(streamlit_dir, 'config.toml')
    with open(config_path, 'w') as f:
        f.write('[theme]\n')
        f.write('primaryColor="#0A2C37"\n')
        f.write('backgroundColor="#FFFFFF"\n')
        f.write('secondaryBackgroundColor="F5F5F5"\n')
        f.write('textColor="#0A2C37"\n')
        f.write('font="sans serif"\n')

    cred_path = os.path.join(streamlit_dir, 'credentials.toml')
    with open(cred_path, 'w') as f:
        f.write('[general]\n')
        f.write('email=""\n')

    #run leaderboard with subprocess
    leaderboard_path = pkg_resources.resource_filename(
        'trulens_eval', 'Leaderboard.py'
    )

    env_opts = {}
    if _dev:
        env_opts['env'] = os.environ
        env_opts['env']['PYTHONPATH'] = str(Path.cwd())

    proc = subprocess.Popen(
        ["streamlit", "run", "--server.headless=True", leaderboard_path], **env_opts
    )

    Tru.dashboard_proc = proc

    return proc

run_feedback_functions(record_json, feedback_functions, chain_json=None)

Run a collection of feedback functions and report their result.

Parameters:

Name Type Description Default
record_json JSON

The record on which to evaluate the feedback

required
chain_json JSON

The chain that produced the given record.

None
feedback_functions Sequence[Feedback]

A collection of feedback

required

Returns nothing.

Source code in trulens_eval/trulens_eval/tru.py
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
def run_feedback_functions(
    self,
    record_json: JSON,
    feedback_functions: Sequence['Feedback'],
    chain_json: Optional[JSON] = None,
) -> Sequence[JSON]:
    """
    Run a collection of feedback functions and report their result.

    Parameters:

        record_json (JSON): The record on which to evaluate the feedback
        functions.

        chain_json (JSON, optional): The chain that produced the given record.
        If not provided, it is looked up from the given database `db`.

        feedback_functions (Sequence[Feedback]): A collection of feedback
        functions to evaluate.

    Returns nothing.
    """

    chain_id = record_json['chain_id']

    if chain_json is None:
        chain_json = self.db.get_chain(chain_id=chain_id)
        if chain_json is None:
            raise RuntimeError(
                "Chain {chain_id} not present in db. "
                "Either add it with `tru.add_chain` or provide `chain_json` to `tru.run_feedback_functions`."
            )

    else:
        assert chain_id == chain_json[
            'chain_id'], "Record was produced by a different chain."

        if self.db.get_chain(chain_id=chain_json['chain_id']) is None:
            logging.warn(
                "Chain {chain_id} was not present in database. Adding it."
            )
            self.add_chain(chain_json=chain_json)

    evals = []

    for func in feedback_functions:
        evals.append(
            TP().promise(
                lambda f: f.run_on_record(
                    chain_json=chain_json, record_json=record_json
                ), func
            )
        )

    evals = map(lambda p: p.get(), evals)

    return list(evals)

start_evaluator(fork=False)

Start a deferred feedback function evaluation thread.

Source code in trulens_eval/trulens_eval/tru.py
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
def start_evaluator(self, fork=False) -> Union[Process, Thread]:
    """
    Start a deferred feedback function evaluation thread.
    """

    assert not fork, "Fork mode not yet implemented."

    if self.evaluator_proc is not None:
        raise RuntimeError("Evaluator is already running in this process.")

    from trulens_eval.tru_feedback import Feedback

    if not fork:
        self.evaluator_stop = threading.Event()

    def runloop():
        while fork or not self.evaluator_stop.is_set():
            print("Looking for things to do. Stop me with `tru.stop_evaluator()`.", end='')
            Feedback.evaluate_deferred(tru=self)
            TP().finish(timeout=10)
            if fork:
                sleep(10)
            else:
                self.evaluator_stop.wait(10)

        print("Evaluator stopped.")

    if fork:
        proc = Process(target=runloop)
    else:
        proc = Thread(target=runloop)

    # Start a persistent thread or process that evaluates feedback functions.

    self.evaluator_proc = proc
    proc.start()

    return proc

stop_dashboard()

Stop existing dashboard if running.

Raises:

Type Description
ValueError

Dashboard is already running.

Source code in trulens_eval/trulens_eval/tru.py
282
283
284
285
286
287
288
289
290
291
292
def stop_dashboard(self) -> None:
    """Stop existing dashboard if running.

    Raises:
        ValueError: Dashboard is already running.
    """
    if Tru.dashboard_proc is None:
        raise ValueError("Dashboard not running.")

    Tru.dashboard_proc.kill()
    Tru.dashboard_proc = None

stop_evaluator()

Stop the deferred feedback evaluation thread.

Source code in trulens_eval/trulens_eval/tru.py
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
def stop_evaluator(self):
    """
    Stop the deferred feedback evaluation thread.
    """

    if self.evaluator_proc is None:
        raise RuntimeError("Evaluator not running this process.")

    if isinstance(self.evaluator_proc, Process):
        self.evaluator_proc.terminate()

    elif isinstance(self.evaluator_proc, Thread):
        self.evaluator_stop.set()
        self.evaluator_proc.join()
        self.evaluator_stop = None

    self.evaluator_proc = None