Mercury

Mercury is a semantic-assisted, cross-text text labeling tool.

1. semantic-assisted: when you select a text span, semantically related text segments will be highlighted -- so you don't have to eyebal through lengthy texts.
2. cross-text: you are labeling text spans from two different texts.

Therefore, Mercury is very efficient for the labeling of NLP tasks that involve comparing texts between two documents which are also lengthy, such as hallucination detection or factual consistency/faithfulness in RAG systems. Semantic assistance not only saves time and reduces fatigues but also avoids mistakes.

Currently, Mercury only supports labeling inconsistencies between the source and summary for summarization in RAG.

Dependencies

Note

You need Python and Node.js.

Mercury uses sqlite-vec to store and search embeddings.

1. pip3 install -r requirements.txt && python3 -m spacy download en_core_web_sm

2. If you don't have pnpm installed: npm install -g pnpm, you may need sudo. If you don't have npm, try sudo apt install npm.

3. To use sqlite-vec via Python's built-in sqlite3 module, you must have SQLite>3.41 (otherwise LIMIT or k=? will not work properly with rowid IN (?) for vector search) installed and set Python's built-in sqlite3 module to use it. Python's built-in sqlite3 module uses its own binary library that is independent of the OS's SQLite. So upgrading the OS's SQLite will not affect Python's sqlite3 module. You need to follow the steps below:

   • Download and compile SQLite>3.41.0 from source

     wget https://www.sqlite.org/2024/sqlite-autoconf-3460100.tar.gz
     tar -xvf sqlite-autoconf-3460100.tar.gz
     cd sqlite-autoconf-3460100
     ./configure
     make

   • Set Python's built-in sqlite3 module to use the compiled SQLite. Suppose you are currently at path $SQLITE_Compile. Then set this environment variable (feel free to replace $SQLITE_Compile with the actual absolute/relative path):

     export LD_PRELAOD=$SQLITE_Compile/.libs/libsqlite3.so

     You may add the above line to ~.bashrc to make it permanent.

   • Verify that Python's sqlite3 module is using the correct SQLite, run this Python code:

     python3 -c "import sqlite3; print(sqlite3.sqlite_version)"

     If the output is the version of SQLite you just compiled, you are good to go.

   • If you are using Mac and run into troubles, please follow SQLite-vec's instructions.

4. To use sqlite-vec directly in sqlite prompt, simply compile sqlite-vec from source and load the compiled vec0.o. The usage can be found in the README of SQLite-vec.

Usage

1. Ingest data for labeling

   Run python3 ingester.py -h to see the options.

   The ingester takes a CSV, JSON, or JSONL file and loads texts from two text columns (configurable via option ingest_column_1 and ingest_column_2 which default to source and summary) of the file. Mercury uses three Vectara corpora to store the sources, the summaries, and the human annotations. You can provide the corpus IDs to overwrite or append data to existing corpora.

2. pnpm install && pnpm build (You need to recompile the frontend each time the UI code changes.)

3. Manually set the labels for annotators to choose from in the labels.yaml file. Mercury supports hierarchical labels.

4. python3 server.py. Be sure to set the candidate labels to choose from in the server.py file.

The annotations are stored in the annotations table in a SQLite database (hardcoded name mercury.sqlite). See the section annotations table for the schema.

The dumped human annotations are stored in a JSON format like this:

[
    {# first sample 
        'sample_id': int,
        'source': str, 
        'summary': str,
        'annotations': [ # a list of annotations from many human annotators
            {
                'annot_id': int,
                'sample_id': int, # relative to the ingestion file
                'annotator': str,  # the annotator unique id
                'annotator_name': str, # the annotator name
                'label': list[str],
                'note': str,
                'summary_span': str, # the text span in the summary
                'summary_start': int,
                'summary_end': int,
                'source_span': str, # the text span in the source
                'source_start': int,
                'source_end': int,
            }
        ], 
        'meta_field_1': Any, # whatever meta info about the sample
        'meta_field_2': Any, 
        ...
    }, 
    {# second sample
        ...
    }, 
    ...
]

You can view exported data in http://[your_host]/viewer

Technical details

Terminology:

• A sample is a pair of source and summary.
• A document is either a source or a summary.
• A chunk is a sentence in a document.

[!NOTE] SQLite uses 1-indexed for autoincrement columns while the rest of the code uses 0-indexed.

Tables

Three tables: chunks, embeddings, annotations, users and leaderboard. All powered by SQLite. In particular, embeddings is powered by sqlite-vec.

chunks table: chunks and metadata

Each row is a chunk.

A JSONL file like this:

# test.jsonl
{"source": "The quick brown fox. Jumps over a lazy dog. ", "summary": "26 letters."}
{"source": "We the people. Of the U.S.A. ", "summary": "The U.S. Constitution. It is great. "}

will be ingested into the chunks table as below:

chunk_id	text	text_type	sample _id	char _offset	chunk _offset
0	"The quick brown fox."	source	0	0	0
1	"Jumps over the lazy dog."	source	0	21	1
2	"We the people."	source	1	0	0
3	"Of the U.S.A."	source	1	15	1
4	"26 letters."	summary	0	0	0
5	"The U.S. Constitution."	summary	1	0	0
6	"It is great."	summary	1	23	1

Meaning of select columns:

• char_offset is the offset of a chunk in its parent document measured by the starting character of the chunk. It allows us to find the chunk in the document.
• chunk_offset_local is the index of a chunk in its parent document. It is used to find the chunk in the document.
• text_type is takes value from the ingestion file. source and summary for now.
• All columns are 0-indexed.
• The sample_id is the index of the sample in the ingestion file. Because the ingestion file could be randomly sampled from a bigger dataset, the sample_id is not necessarily global.

embeddings table: the embeddings of chunks

rowid	embedding
1	[0.1, 0.2, ..., 0.9]
2	[0.2, 0.3, ..., 0.8]

• rowid here and chunk_id in the chunks table have one-to-one correspondence. rowid is 1-indexed due to sqlite-vec. We cannot do anything about it. So when aligning the tables chunks and embeddings, remember to subtract 1 from rowid to get chunk_id.

annotations table: the human annotations

annot_id	sample _id	annot_spans	annotator	label	note
1	1	{'source': [1, 10], 'summary': [7, 10]}	2fe9bb69	["ambivalent"]	"I am not sure."
2	1	{'summary': [2, 8]}	a24cb15c	["extrinsic"]	"No connection to the source."

• sample_id are the id's of chunks in the chunks table.
• text_spans is a JSON text field that stores the text spans selected by the annotator. Each entry is a dictionary where keys must be those in the text_type column in the chunks table (hardcoded to source and summary now) and the values are lists of two integers: the start and end indices of the text span in the chunk. For extrinsic hallucinations (no connection to the source at all), only summary-key items. The reason we use JSON here is that SQLite does not support array types.

config table: the configuration

For example:

key	value
embdding_model	"openai/text-embedding-3-small"
embdding_dimension	4

sample_meta table: the sample metadata

sample_id	json_meta
0	{"model":"meta-llama/Meta-Llama-3.1-70B-Instruct","HHEMv1":0.43335,"HHEM-2.1":0.39717,"HHEM-2.1-English":0.90258,"trueteacher":1,"true_nli":0.0,"gpt-3.5-turbo":1,"gpt-4-turbo":1,"gpt-4o":1, "sample_id":727}
1	{"model":"openai/GPT-3.5-Turbo","HHEMv1":0.43003,"HHEM-2.1":0.97216,"HHEM-2.1-English":0.92742,"trueteacher":1,"true_nli":1.0,"gpt-3.5-turbo":1,"gpt-4-turbo":1,"gpt-4o":1, "sample_id": 1018}

0-indexed, the sample_id column is the sample_id in the chunks table. It is local to the ingestion file. The json_meta is whatever info other than ingestion columns (source and summary) in the ingestion file.

users table: the annotators

user_id	user_name
add93a266ab7484abdc623ddc3bf6441	Alice
68d41e465458473c8ca1959614093da7	Bob

How to do vector search

SQLite-vec uses Euclidean distance for vector search. So all embeddings much be normalized to unit length. Fortunately, OpenAI and Sentence-Bert's embeddings are already normalized.

1. Suppose the user selects a text span in chunk of global chunk ID x. Assume that the text span selection cannot cross sentence boundaries.
2. Get x's doc_id from the chunks table.
3. Get x's embedding from the embeddings table by where rowid = {chunk_id}. Denote it as x_embedding.
4. Get the chunk_ids of all chunks in the opposite document (source if x is in summary, and vice versa) by where doc_id = {doc_id} and text_type={text_type}. Denote such chunk IDs as y1, y2, ..., yn.
5. Send a query to SQLite like this:

     SELECT
        rowid,
        distance
      FROM embeddings
      WHERE embedding MATCH '{x_embedding}'
      and rowid in ({y1, y2, ..., yn}) 
      ORDER BY distance 
      LIMIT 5

   This will find the 5 most similar chunks to x in the opposite document. It limits vector search within the opposite document by rowid in (y1, y2, ..., yn). Note that rowid, embedding, and distance are predefined by sqlite-vec.

Here is a running example (using the data above):

1. Suppose the data has been ingested. The embedder is openai/text-embedding-3-small` and the embedding dimension is 4.
2. Suppose the user selects sample_id = 1 and chunk_id = 5: "The U.S. Constitution." The text_type of chunk_id = 5 is summary -- the opposite document is the source.
3. Let's get the chunk IDs of the source document:

   SELECT chunk_id
   FROM chunks
   WHERE sample_id = 1 and text_type = 'source'

   The return is 2, 3.
4. The embedding of "The U.S. Constitution" can be obtained from the embeddings table by where rowid = 6. Note that because SQLite uses 1-indexed, so we need to add 1 from chunk_id to get rowid.

    SELECT embedding
    FROM embeddings
    WHERE rowid = 6

   The return is [0.08553484082221985, 0.21519172191619873, 0.46908700466156006, 0.8522521257400513].
5. Now We search for its nearest neighbors in its corresponding source chunks of rowid 4 and 5 -- again, obtained by adding 1 from chunk_id 2 and 3 obtained in step 3.

   SELECT
       rowid,
       distance
   FROM embeddings
   WHERE embedding MATCH '[0.08553484082221985, 0.21519172191619873, 0.46908700466156006, 0.8522521257400513]'
   and rowid in (4, 5) 
   ORDER BY distance

   The return is [(4, 0.3506483733654022), (5, 1.1732779741287231)].
6. Translate the rowid back to chunk_id by subtracting 4 and 5 to get 2 and 3. The closest source chunk is "We the people" (rowid=3 while chunk_id=2) which is the most famous three words in the US Constitution.

Limitations

1. OpenAI's embedding endpoint can only embed up to 8192 tokens in each call.
2. embdding_dimension is only useful for OpenAI models. Most other models do not support changing the embedding dimension.

Embedding speed and/or embedding dimension

1. multi-qa-mpnet-base-dot-v1 takes about 0.219 second on a x86 CPU to embed one sentence when batch_size is 1. The embedding dimension is 768.
2. BAAI/bge-small-en-v1.5 takes also about 0.202 second on a x86 CPU to embed one sentence when batch_size is 1. The embedding dimension is 384.