Agentic Multi-Stage RAG with Cohere Tools API

1
import os
2
from pprint import pprint
3
4
import cohere
5
import pandas as pd
6
from sklearn.metrics.pairwise import cosine_similarity

1
# versions
2
print('cohere version:', cohere.__version__)

cohere version: 5.5.1

1
COHERE_API_KEY = os.environ.get("CO_API_KEY")
2
COHERE_MODEL = 'command-r-plus'
3
co = cohere.Client(api_key=COHERE_API_KEY)

1
documents = [
2
    {
3
        "title": "Bicycle law",
4
        "body": """
5
        Traffic Infractions and fees - For all information related to bicycle traffic infractions such as not wearing a helmet and fee information, please visit Section 3b for more information.
6
        Riding on the road - When riding on a roadway, a cyclist has all the rights and responsibilities of a vehicle driver (RCW 46.61.755). Bicyclists who violate traffic laws may be ticketed (RCW 46.61.750).
7
        Roads closed to bicyclists - Some designated sections of the state's limited access highway system may be closed to bicyclists. See the permanent bike restrictions map for more information. In addition, local governments may adopt ordinances banning cycling on specific roads or on sidewalks within business districts.
8
        Children bicycling - Parents or guardians may not knowingly permit bicycle traffic violations by their ward (RCW 46.61.700).
9
        Riding side by side - Bicyclists may ride side by side, but not more than two abreast (RCW 46.61.770).
10
        Riding at night - For night bicycle riding, a white front light (not a reflector) visible for 500 feet and a red rear reflector are required. A red rear light may be used in addition to the required reflector (RCW 46.61.780).
11
        Shoulder vs. bike lane - Bicyclists may choose to ride on the path, bike lane, shoulder or travel lane as suits their safety needs (RCW 46.61.770).
12
        Bicycle helmets - Currently, there is no state law requiring helmet use. However, some cities and counties do require helmets. For specific information along with location for bicycle helmet law please reference to section 21a.
13
        Bicycle equipment - Bicycles must be equipped with a white front light visible for 500 feet and a red rear reflector (RCW 46.61.780). A red rear light may be used in addition to the required reflector.
14
""",
15
    },
16
    {
17
        "title": "Bicycle helmet requirement",
18
        "body": "Currently, there is no state law requiring helmet use. However, some cities and counties do require helmet use with bicycles. Here is a list of those locations and when the laws were enacted. For specific information along with location for bicycle helmet law please reference to section 21a.",
19
    },
20
    {
21
        "title": "Section 21a",
22
        "body": """helmet rules by location: These are city and county level rules. The following group must wear helmets.
23
        Location name | Who is affected | Effective date
24
        Aberdeen | All ages | 2001
25
        Bainbridge Island | All ages | 2001
26
        Bellevue | All ages | 2001
27
        Bremerton | All ages | 2000
28
        DuPont | All ages | 2008
29
        Eatonville | All ages | 1996
30
        Fircrest | All ages | 1995
31
        Gig Harbor | All ages | 1996
32
        Kent | All ages | 1999
33
        Lynnwood | All ages | 2004
34
        Lakewood | All ages | 1996
35
        Milton | All ages | 1997
36
        Orting | Under 17 | 1997
37
38
     For fines and rules, you will be charged in according with Section 3b of the law.
39
     """,
40
    },
41
    {
42
        "title": "Section 3b",
43
        "body": """Traffic infraction - A person operating a bicycle upon a roadway or highway shall be subject to the provisions of this chapter relating to traffic infractions.
44
        1. Stop for people in crosswalks. Every intersection is a crosswalk - It’s the law. Drivers must stop for pedestrians at intersections, whether it’s an unmarked or marked crosswalk, and bicyclists in crosswalks are considered pedestrians. Also, it is illegal to pass another vehicle stopped for someone at a crosswalk. In Washington, the leading action motorists take that results in them hitting someone is a failure to yield to pedestrians.
45
        2. Put the phone down. Hand-held cell phone use and texting is prohibited for all Washington drivers and may result in a $136 fine for first offense, $235 on the second distracted-driving citation.
46
        3. Helmets are required for all bicyclists according to the state and municipal laws. If you are in a group required to wear a helmet but do not wear it you can be fined $48. # If you are the parent or legal guardian of a child under 17 and knowingly allow them to ride without a helmet, you can be fined $136.
47
""",
48
    },
49
]
50
db = pd.DataFrame(documents)
51
# comebine title and body
52
db["combined"] = "Title: " + db["title"] + "\n" + "Body: " + db["body"]
53
# generate embedding
54
embeddings = co.embed(
55
    texts=db.combined.tolist(), model="embed-english-v3.0", input_type="search_document"
56
)
57
db["embeddings"] = embeddings.embeddings

1
db

1
def retrieve_documents(query: str, n=1) -> dict:
2
    """
3
    Function to retrieve documents a given query.
4
5
    Steps:
6
    1. Embed the query
7
    2. Calculate cosine similarity between the query embedding and the embeddings of the documents
8
    3. Return the top n documents with the highest similarity scores
9
    """
10
    query_emb = co.embed(
11
        texts=[query], model="embed-english-v3.0", input_type="search_query"
12
    )
13
14
    similarity_scores = cosine_similarity(
15
        [query_emb.embeddings[0]], db.embeddings.tolist()
16
    )
17
    similarity_scores = similarity_scores[0]
18
19
    top_indices = similarity_scores.argsort()[::-1][:n]
20
    top_matches = db.iloc[top_indices]
21
22
    return {"top_matched_document": top_matches.combined}
23
24
25
functions_map = {
26
    "retrieve_documents": retrieve_documents,
27
}
28
29
tools = [
30
    {
31
        "name": "retrieve_documents",
32
        "description": "given a query, retrieve documents from a database to answer user's question",
33
        "parameter_definitions": {
34
            "query": {"description": "query", "type": "str", "required": True}
35
        },
36
    }
37
]

1
def simple_rag(query, db):
2
    """
3
    Given user's query, retrieve top documents and generate response using documents parameter.
4
    """
5
    top_matched_document = retrieve_documents(query)["top_matched_document"]
6
7
    print("top_matched_document", top_matched_document)
8
9
    output = co.chat(
10
        message=query, model=COHERE_MODEL, documents=[top_matched_document]
11
    )
12
13
    return output.text

1
def cohere_agent(
2
    message: str,
3
    preamble: str,
4
    tools: list[dict],
5
    force_single_step=False,
6
    verbose: bool = False,
7
    temperature: float = 0.3,
8
) -> str:
9
    """
10
    Function to handle multi-step tool use api.
11
12
    Args:
13
        message (str): The message to send to the Cohere AI model.
14
        preamble (str): The preamble or context for the conversation.
15
        tools (list of dict): List of tools to use in the conversation.
16
        verbose (bool, optional): Whether to print verbose output. Defaults to False.
17
18
    Returns:
19
        str: The final response from the call.
20
    """
21
22
    counter = 1
23
24
    response = co.chat(
25
        model=COHERE_MODEL,
26
        message=message,
27
        preamble=preamble,
28
        tools=tools,
29
        force_single_step=force_single_step,
30
        temperature=temperature,
31
    )
32
33
    if verbose:
34
        print(f"\nrunning 0th step.")
35
        print(response.text)
36
37
    while response.tool_calls:
38
        tool_results = []
39
40
        if verbose:
41
            print(f"\nrunning {counter}th step.")
42
43
        for tool_call in response.tool_calls:
44
            output = functions_map[tool_call.name](**tool_call.parameters)
45
            outputs = [output]
46
            tool_results.append({"call": tool_call, "outputs": outputs})
47
48
            if verbose:
49
                print(
50
                    f"= running tool {tool_call.name}, with parameters: \n{tool_call.parameters}"
51
                )
52
                print(f"== tool results:")
53
                pprint(output)
54
55
        response = co.chat(
56
            model=COHERE_MODEL,
57
            message="",
58
            chat_history=response.chat_history,
59
            preamble=preamble,
60
            tools=tools,
61
            force_single_step=force_single_step,
62
            tool_results=tool_results,
63
            temperature=temperature,
64
        )
65
66
        if verbose:
67
            print(response.text)
68
            counter += 1
69
70
    return response.text

1
question1 = "Is there a state level law for wearing helmets?"

1
output = simple_rag(question1, db)
2
print(output)

top_matched_document 1    Title: Bicycle helmet requirement\nBody: Curre...
Name: combined, dtype: object
There is currently no state law requiring the use of helmets when riding a bicycle. However, some cities and counties do require helmet use.

1
preamble = """
2
You are an expert assistant that helps users answers question about legal documents and policies.
3
Use the provided documents to answer questions about an employee's specific situation.
4
"""
5
6
output = cohere_agent(question1, preamble, tools, verbose=True)

running 0th step.
I will search for 'state level law for wearing helmets' in the documents provided and write an answer based on what I find.
running 1th step.
= running tool retrieve_documents, with parameters:
{'query': 'state level law for wearing helmets'}
== tool results:
{'top_matched_document': 1    Title: Bicycle helmet requirement\nBody: Curre...
Name: combined, dtype: object}
There is currently no state law requiring helmet use. However, some cities and counties do require helmet use with bicycles.

1
question2 = "I live in orting, do I need to wear a helmet with a bike?"

1
output = simple_rag(question2, db)
2
print(output)

top_matched_document 1    Title: Bicycle helmet requirement\nBody: Curre...
Name: combined, dtype: object
In the state of Washington, there is no law requiring you to wear a helmet when riding a bike. However, some cities and counties do require helmet use, so it is worth checking your local laws.

1
preamble = """
2
You are an expert assistant that helps users answers question about legal documents and policies.
3
Use the provided documents to answer questions about an employee's specific situation.
4
"""
5
6
output = cohere_agent(question2, preamble, tools, verbose=True)

running 0th step.
I will search for 'helmet with a bike' and then write an answer.
running 1th step.
= running tool retrieve_documents, with parameters:
{'query': 'helmet with a bike'}
== tool results:
{'top_matched_document': 1    Title: Bicycle helmet requirement\nBody: Curre...
Name: combined, dtype: object}
There is no state law requiring helmet use, however, some cities and counties do require helmet use with bicycles. I cannot find any information about Orting specifically, but you should check with your local authority.

1
def reference_extractor(query: str, documents: list[str]) -> str:
2
    """
3
    Given a query and document, find references to other documents.
4
    """
5
    prompt = f"""
6
    # instruction
7
    Does the reference document mention any other documents? If so, list them.
8
    If not, return empty string.
9
10
    # user query
11
    {query}
12
13
    # retrieved documents
14
    {documents}
15
    """
16
17
    return co.chat(message=prompt, model=COHERE_MODEL, preamble=None).text
18
19
20
def retrieve_documents(query: str, n=1) -> dict:
21
    """
22
    Function to retrieve most relevant documents a given query.
23
    It also returns other references mentioned in the top matched documents.
24
    """
25
    query_emb = co.embed(
26
        texts=[query], model="embed-english-v3.0", input_type="search_query"
27
    )
28
29
    similarity_scores = cosine_similarity(
30
        [query_emb.embeddings[0]], db.embeddings.tolist()
31
    )
32
    similarity_scores = similarity_scores[0]
33
34
    top_indices = similarity_scores.argsort()[::-1][:n]
35
    top_matches = db.iloc[top_indices]
36
    other_references = reference_extractor(query, top_matches.combined.tolist())
37
38
    return {
39
        "top_matched_document": top_matches.combined,
40
        "other_references_to_query": other_references,
41
    }
42
43
44
functions_map = {
45
    "retrieve_documents": retrieve_documents,
46
}
47
48
tools = [
49
    {
50
        "name": "retrieve_documents",
51
        "description": "given a query, retrieve documents from a database to answer user's question. It also finds references to other documents that should be leveraged to retrieve more documents",
52
        "parameter_definitions": {
53
            "query": {
54
                "description": "user's question or question or name of other document sections or references.",
55
                "type": "str",
56
                "required": True,
57
            }
58
        },
59
    }
60
]

1
preamble2 = """# Instruction
2
You are an expert assistant that helps users answers question about legal documents and policies.
3
4
Please follow these steps:
5
1. Using user's query, use `retrieve_documents` tool to retrieve the most relevant document from the database.
6
2. If you see `other_references_to_query` in the tool result, search the mentioned referenced using `retrieve_documents(<other reference="">)` tool to retrieve more documents.
7
3. Keep trying until you find the answer.
8
4. Answer with yes or no as much as you can to answer the question directly.
9
"""
10
11
output = cohere_agent(question2, preamble2, tools, verbose=True)

running 0th step.
I will search for 'Orting' and 'bike helmet' to find the relevant information.
running 1th step.
= running tool retrieve_documents, with parameters:
{'query': 'Orting bike helmet'}
== tool results:
{'other_references_to_query': 'Section 21a, Section 3b',
    'top_matched_document': 0    Title: Bicycle law\nBody: \n        Riding on ...
Name: combined, dtype: object}
I have found that there is no state law requiring helmet use, but some cities and counties do require helmets. I will now search for 'Section 21a' to find out if Orting is one of these cities or counties.
running 2th step.
= running tool retrieve_documents, with parameters:
{'query': 'Section 21a'}
== tool results:
{'other_references_to_query': '- Section 3b',
    'top_matched_document': 2    Title: Section 21a\nBody: helmet rules by loca...
Name: combined, dtype: object}
Yes, you do need to wear a helmet when riding a bike in Orting if you are under 17.