Government Funding Graph RAG | Towards Data Science

, I current my newest open-source challenge — Authorities Funding Graph.

The inspiration for this challenge got here from a want to make higher tooling for grant writing, specifically to counsel analysis matters, funding our bodies, analysis establishments, and researchers. I’ve made Innovate UK grant purposes up to now, so I’ve had an curiosity within the authorities funding panorama for a while.

Concretely, a whole lot of the current political discourse focuses on authorities spending, specifically Elon Musk’s Department of Government Efficiency (DOGE) in the US and comparable sentiments echoed right here within the UK, as Kier Starmer appears to combine AI into government.

Maybe the discharge of this challenge is kind of well timed. Albeit not the unique intention, I hope as a secondary consequence of this text is that it evokes extra exploration into open supply datasets for public spending.

I’ve used Networkx & PyVis to visualise the graph of UKRI API information. Then, I element a LlamaIndex graph RAG implementation. For completeness, I’ve additionally included my preliminary LangChain-based resolution. The net framework is Streamlit, the demo is hosted on Streamlit group cloud.

This text accommodates the next sections.

1. Definitions
2. UKRI API
3. Assemble NetworkX Graph
4. Filter a NetworkX Graph
5. Graph Visualisation Utilizing PyVis
6. Graph RAG Utilizing LlamaIndex
7. Linting With Pylint
8. Streamlit Group Cloud Demo App (on the very finish of the article)

1. Definitions

What’s UKRI?

UK Analysis and Innovation is a non-departmental public physique sponsored by the Division for Science, Innovation and Know-how (DSIT) that allocates funding for analysis and improvement. Usually, funding is awarded to analysis establishments and companies.

“We make investments £8 billion of taxpayers’ cash annually into analysis and innovation and the individuals who make it occur. We work throughout an enormous vary of fields — from biodiversity conservation to quantum computing, and from house telescopes to modern well being care. We give everybody the chance to contribute and to profit, bringing collectively folks and organisations nationally and globally to create, develop and deploy new concepts and applied sciences.” — UKRI Website

What’s a Graph?

A graph is a handy information construction displaying the relationships between completely different entities (nodes) and their relationships to one another (edges). In some situations, we additionally affiliate these relationships with a numerical worth.

“In laptop science, a graph is an summary information sort that’s meant to implement the undirected graph and directed graph ideas from the sector of graph concept inside arithmetic.

A graph information construction consists of a finite (and probably mutable) set of vertices (additionally known as nodes or factors), along with a set of unordered pairs of those vertices for an undirected graph or a set of ordered pairs for a directed graph. These pairs are often known as edges (additionally known as hyperlinks or strains), and for a directed graph are also called edges but additionally typically arrows or arcs.” — Wikipedia

What’s NetworkX?

NetworkX is a helpful library on this challenge to assemble and retailer our graph. Particularly, a digraph although the library helps many graph variants equivalent to multigraphs, the library additionally helps graph-related utility capabilities.

“NetworkX is a Python bundle for the creation, manipulation, and research of the construction, dynamics, and capabilities of complicated networks.” — NetworkX Website

What’s PyVis?

We use the PyVis Python bundle to create dynamic community views for our graph, screenshots of those may be discovered all through the article.

“The pyvis library is supposed for fast technology of visible community graphs with minimal python code. It’s designed as a wrapper across the well-liked Javascript visJS library” — PyVis Docs

What’s LlamaIndex?

LlamaIndex is a well-liked library for LLM purposes, together with help for agentic workflows, we use it to carry out the graph RAG element of this challenge.

“LlamaIndex (GPT Index) is an information framework in your LLM utility. Constructing with LlamaIndex usually includes working with LlamaIndex core and a selected set of integrations (or plugins).” — LlamaIndex Github

What’s Graph RAG?

Retrieval-augmented technology, or RAG as it’s generally recognized, is an AI framework for which extra context from an exterior information base is used to floor LLM solutions. Graph RAG, by extension, pertains to the usage of a Graph to supply this extra context.

“GraphRAG is a robust retrieval mechanism that improves GenAI purposes by profiting from the wealthy context in graph information buildings… Fundamental RAG programs rely solely on semantic search in vector databases to retrieve and rank units of remoted textual content fragments. Whereas this method can floor some related info, it fails to seize the context connecting these items. Because of this, fundamental RAG programs are ill-equipped to reply complicated, multi-hop questions. That is the place GraphRAG is available in. It makes use of information graphs to symbolize and join info to seize not solely extra information factors but additionally their relationships. Thus, graph-based retrievers can present extra correct and related outcomes by uncovering hidden connections that aren’t typically apparent however are essential for correlating info.” — Neo4j Website

What’s Streamlit?

Streamlit is a light-weight Python internet framework we’ll use to create the online utility for this challenge.

“Streamlit is an open-source Python framework for information scientists and AI/ML engineers to ship dynamic information apps with only some strains of code. Construct and deploy highly effective information apps in minutes.” — Streamlit website

2. UKRI API

The UKRI API is a service that facilitates entry to the general public UKRI grant funding dataset, authentication isn’t required and the docs may be discovered here. I exploit solely two endpoints for our utility, they’re the Search projects endpoint and the Projects endpoint. This enables a consumer to seek for tasks based mostly on a key phrase search and retrieve all project-specific info.

A search time period, web page measurement and web page quantity are supplied as question string parameters. The question string parameters;

selectedSortableField=professional.am&selectedSortOrder=DESC

Make sure that the outcomes are returned by funded worth descending.

I’ve additionally included the code I used for asynchronous pagination.

import math
import requests
import concurrent.futures
import os 
from itertools import chain
import urllib.parse
import logging

def search_ukri_projects(args):
    """
    Search UKRI tasks based mostly on a search time period web page measurement and web page quantity.
    Extra particulars may be discovered right here: https://gtr.ukri.org/sources/api.html
    """
    search_term, page_size, page_number = args
    strive:
        encoded_search_term = urllib.parse.quote(search_term)
        if (
            (
                response := requests.get(
                    f"https://gtr.ukri.org/api/search/challenge?time period={encoded_search_term}&web page={page_number}&fetchSize={page_size}&selectedSortableField=professional.am&selectedSortOrder=DESC&selectedFacets=&fields=challenge.abs",
                    timeout=10,
                )
            )
            and (response.status_code == 200)
            and (
                objects := response.json()
                .get("facetedSearchResultBean", {})
                .get("outcomes")
            )
        ):
            return objects
    besides Exception as error:
        logging.exception("ERROR search_ukri_projects: %s", error)
    return []

def search_ukri_paginate(search_term, number_of_results, page_size=100):
    """
    Asynchronous pagination requests for challenge lookup.
    """
    args = [
        (search_term, page_size, page_number + 1)
        for page_number in range(int(math.ceil(number_of_results / page_size)))
    ]
    with concurrent.futures.ThreadPoolExecutor(os.cpu_count()) as executor:
        future = executor.map(search_ukri_projects, args)
    outcomes = [result for result in future if result]
    return record(chain.from_iterable(outcomes))[:number_of_results]

The next operate is used to get project-specific information utilizing the distinctive UKRI challenge reference. The challenge reference is derived from the aforementioned challenge search outcomes.

import requests 
import logging

def get_ukri_project_data(project_grant_reference):
    """
    Search UKRI challenge information based mostly on grant reference.
    """
    strive:
        if (
            (
                response := requests.get(
                    f"https://gtr.ukri.org/api/tasks?ref={project_grant_reference}",
                    timeout=10,
                )
            )
            and (response.status_code == 200)
            and (objects := response.json().get("projectOverview", {}))
        ):
            return objects
    besides Exception as error:
        logging.exception("ERROR get_ukri_project_data: %s", error)

Equally, we parse out the related information for the development of the graph and take away superfluous info.

def parse_data(tasks):
    """
    Parse challenge information right into a usable format and validate.
    """
    information = []
    for challenge in tasks:
        project_composition = challenge.get("projectComposition", {})
        project_data = project_composition.get("challenge", {})
        fund = project_data.get("fund", {})
        funder = fund.get("funder")
        value_pounds = fund.get("valuePounds")
        lead_research_organisation = project_composition.get("leadResearchOrganisation")
        person_roles = project_composition.get("personRoles")
        if all(
            [
                project_composition,
                project_data,
                fund,
                funder,
                value_pounds,
                lead_research_organisation,
            ]
        ):
            document = {}
            document["funder_name"] = funder.get("title")
            document["funder_link"] = funder.get("resourceUrl")
            document["project_title"] = project_data.get("title")
            document["project_grant_reference"] = project_data.get("grantReference")
            document["value"] = value_pounds
            document["lead_research_organisation"] = lead_research_organisation.get(
                "title", ""
            )
            document["lead_research_organisation_link"] = lead_research_organisation.get(
                "resourceUrl", ""
            )
            document["people"] = person_roles
            document["project_url"] = project_data.get("resourceUrl")
            information.append(document)
    return information

3. Assemble NetworkX Graph

There are several types of graphs, and I elected for a directed graph the place the route of the sides are vital. Extra formally;

“A DiGraph shops nodes and edges with non-obligatory information, or attributes. DiGraphs maintain directed edges. Self loops are allowed however a number of (parallel) edges are usually not.” — NetworkX Website

To assemble the NetworkX graph, we should add nodes and edges — together with the sequential updating of node attributes.

The usual attributes, suitable with PyVis graph rendering for nodes are as follows;

• Title (The label that seems on hover over)
• Group (The color coding)
• Measurement (How massive the nodes seem within the graph)

We additionally use the customized attribute “funding”, which we’ll use to sum the entire funding for analysis and funding organizations. This shall be normalized to set the node measurement in keeping with the proportion of whole funding for a specific group.

For our graph, we’ve nodes from 4 teams. They’re categorised as: funder_name, lead_research_organisation, project_title and person_name.

HTML hyperlinks can be utilized within the node title to permit the consumer to simply click on via to a URL. I’ve included a helper operate to do that beneath. There are challenge, folks, and analysis organisation-specific hyperlinks that, if redirected to supply extra info to the consumer.

The code to assemble the NetworkX graph may be seen beneath. The DiGraph class has strategies to verify if a graph already has a node and equally for edges. There are additionally strategies for including nodes and edges. As we iterate via tasks, we wish to sum the entire funding quantity for the funding group and lead analysis establishment. There are strategies to each get an attribute from a node within the graph and set an attribute on a node. Relying on the supply and vacation spot node, we additionally apply completely different titles and labels to replicate that particular predicate. These may be seen within the code beneath.

import networkx as nx

def get_link_html(hyperlink, textual content):
    """
    Helper operate to assemble a HTML hyperlink.
    """
    return f"""{text}"""

def set_networkx_attribute(graph, node_label, attribute_name, worth):
    """
    Helper to set attribute for networkx graph.
    """
    attrs = {node_label: {attribute_name: worth}}
    nx.set_node_attributes(graph, attrs)

def append_networkx_value(graph, node_label, attribute_name, worth):
    """
    Helper to append worth to present node attribute scalar worth.
    """
    current_map = nx.get_node_attributes(graph, attribute_name, default=0)
    current_value = current_map[node_label]
    current_value = current_value + worth
    set_networkx_attribute(graph, node_label, attribute_name, current_value)

def create_networkx(information):
    """
    Create networkx graph from UKRI information.
    """
    graph = nx.DiGraph()
    for row in information:
        if (
            (funder_name := row.get("funder_name"))
            and (project_title := row.get("project_title"))
            and (lead_research_organisation := row.get("lead_research_organisation"))
        ):

            project_data_lookup = row.get("project_data_lookup", {})

            if not graph.has_node(funder_name):
                graph.add_node(
                    funder_name, title=funder_name, group="funder_name", measurement=100
                )
            if not graph.has_node(project_title):
                link_html = get_link_html(
                    row.get("project_url", "").exchange("api/", ""), project_title
                )
                graph.add_node(
                    project_title,
                    title=link_html,
                    group="project_title",
                    project_data_lookup=project_data_lookup,
                    measurement=25,
                )
            if not graph.has_edge(funder_name, project_title):
                graph.add_edge(
                    funder_name,
                    project_title,
                    worth=row.get("worth"),
                    title=f"{'£{:,.2f}'.format(row.get('worth'))}",
                    label=f"{'£{:,.2f}'.format(row.get('worth'))}",
                )

        if not graph.has_node(lead_research_organisation):
            link_html = get_link_html(
                row.get("lead_research_organisation_link").exchange("api/", ""),
                lead_research_organisation,
            )
            graph.add_node(
                lead_research_organisation,
                title=link_html,
                group="lead_research_organisation",
                measurement=50,
            )
        if not graph.has_edge(lead_research_organisation, project_title):
            graph.add_edge(
                lead_research_organisation, project_title, title="RELATES TO"
            )

        append_networkx_value(graph, funder_name, "funding", row.get("worth", 0))
        append_networkx_value(graph, project_title, "funding", row.get("worth", 0))
        append_networkx_value(
            graph, lead_research_organisation, "funding", row.get("worth", 0)
        )

        person_roles = row.get(
            "folks", []
        )  

        for particular person in person_roles:
            if (
                (person_name := particular person.get("fullName"))
                and (person_link := particular person.get("resourceUrl"))
                and (project_title := row.get("project_title"))
                and (roles := particular person.get("roles"))
            ):
                if not graph.has_node(person_name):
                    link_html = get_link_html(
                        person_link.exchange("api/", ""), person_name
                    )
                    graph.add_node(
                        person_name, title=link_html, group="person_name", measurement=10
                    )
                for function in roles:
                    if (not graph.has_edge(person_name, project_title)) or (
                        not graph[person_name][project_title]["title"]
                        == function.get("title")
                    ):
                        graph.add_edge(
                            person_name,
                            project_title,
                            title=function.get("title"),
                            label=function.get("title"),
                        )
    return graph

As soon as the graph has been constructed and as beforehand described, I needed to normalize the node sizes relying on the proportion of the entire quantity of funding for specific teams. I additionally append the entire funding, each as a summation and as a proportion to the node label so it may be extra simply considered by a consumer.

The dimensions issue is only a a number of utilized for aesthetic causes, such that the node sizes seem relative to the opposite node teams current.

import networkx as nx
import math 
import utils.config as config  # pylint: disable=consider-using-from-import, import-error

def set_networkx_attribute(graph, node_label, attribute_name, worth):
    """
    Helper to set attribute for networkx graph.
    """
    attrs = {node_label: {attribute_name: worth}}
    nx.set_node_attributes(graph, attrs)

def calculate_total_funding_from_group(graph, group):
    """
    Helper to calculate whole funding for a bunch.
    """
    return sum(
        [
            data.get("funding")
            for node_label, data in graph.nodes(data=True)
            if data.get("funding") and data.get("group") == group
        ]
    )

def set_weighted_size_helper(graph, node_label, totals, information):
    """
    Create normalized weights based mostly on proportion funding quantity.
    """
    if (
        (group := information.get("group"))
        and (total_funding := totals.get(group))
        and (funding := information.get("funding"))
    ):
        div = funding / total_funding
        funding_percentage = math.ceil(((100.0 * div)))
        set_networkx_attribute(graph, node_label, "measurement", funding_percentage)

def annotate_value_on_graph(graph):
    """
    Calculate normalized graph sizes and append to title.
    """
    totals = {}
    for group in ["lead_research_organisation", "funder_name"]:
        totals[group] = calculate_total_funding_from_group(graph, group)

    for node_label, information in graph.nodes(information=True):
        if (
            (funding := information.get("funding"))
            and (group := information.get("group"))
            and (title := information.get("title"))
        ):
            new_title = f"{title} | {'£ {:,.0f}'.format(funding)}"
            if total_funding := totals.get(group):
                div = funding / total_funding
                funding_percentage = math.ceil(((100.0 * div)))
                set_networkx_attribute(
                    graph,
                    node_label,
                    "measurement",
                    config.NODE_SIZE_SCALE_FACTOR * funding_percentage,
                )
                new_title += f" | {' {:,.0f}'.format(funding_percentage)} %"

            set_networkx_attribute(graph, node_label, "title", new_title)

4. Filter a NetworkX Graph

I permit the consumer to filter nodes through the UI to create a subgraph. The shape to do that in Streamlit is beneath. I additionally discover the neighbors of neighbors for the filtered nodes. I had some points with Pylint elevating pointless comprehension errors from the generator, which I’ve disabled — extra on Pylint later within the article. A smaller graph will take much less time to render and can make sure that irrelevant context shall be excluded.

import networkx as nx
import streamlit as st

def find_neighbor_nodes_helper(node_list, graph):
    """
    Discover distinctive node neighbors and flatten.
    """
    successors_generator_array = [
        # pylint: disable=unnecessary-comprehension
        [item for item in graph.successors(node)]
        for node in node_list
    ]
    predecessors_generator_array = [
        # pylint: disable=unnecessary-comprehension
        [item for item in graph.predecessors(node)]
        for node in node_list
    ]
    neighbors = successors_generator_array + predecessors_generator_array
    flat = sum(neighbors, [])
    return record(set(flat))

def render_filter_form(annotated_node_data, graph):
    """
    Render type to permit the consumer to outline search nodes.
    """
    st.session_state["filter"] = st.radio(
        "Filter", ["No filter", "Filter results"], index=0, horizontal=True
    )
    if (filter_determinant := st.session_state.get("filter")) and (
        filter_determinant == "Filter outcomes"
    ):
        st.session_state["node_group"] = st.selectbox(
            "Entity sort", record(annotated_node_data.keys())
        )
        if node_group := st.session_state.get("node_group"):
            ordered_lookup = dict(
                sorted(
                    annotated_node_data[node_group].objects(),
                    key=lambda merchandise: merchandise[1].get("neighbor_len"),
                    reverse=True,
                )
            )
            st.session_state["search_nodes_label"] = st.multiselect(
                "Filter tasks", record(ordered_lookup.keys())
            )
        if search_nodes_label := st.session_state.get("search_nodes_label"):
            filter_nodes = [
                ordered_lookup[label].get("label") for label in search_nodes_label
            ]
            search_nodes_neighbors = find_neighbor_nodes_helper(filter_nodes, graph)
            search_nodes = find_neighbor_nodes_helper(search_nodes_neighbors, graph)
            st.session_state["search_nodes"] = record(
                set(search_nodes + filter_nodes + search_nodes_neighbors)
            )

NetworkX makes it straightforward to create a subgraph from a listing of nodes with the subgraph_view operate, which takes a callable as a parameter. The callable takes a graph node as a parameter and if the boolean True worth is returned, the node can be included within the subgraph.

import networkx as nx
import streamlit as st

def filter_node(node):
    """
    Examine to see if the filter time period is within the nodes chosen.
    """
    if (
        (filter_term := st.session_state.get("filter"))
        and (filter_term == "Filter outcomes")
        and (search_nodes := st.session_state.get("search_nodes"))
    ):
        if node not in search_nodes:
            return False
    return True

graph = nx.subgraph_view(graph, filter_node=filter_node)

5. Graph Visualisation Utilizing PyVis

To provide the visualizations I’ve introduced earlier within the article, we should first convert the NetworkX graph to a PyVis community after which render the HTML file throughout the Streamlit UI.

If you’re unfamiliar with Streamlit, you may see considered one of my different articles that discover the subject here.

Changing a NetworkX graph to PyVis format is comparatively trivial and may be achieved with the code beneath. The Community class is the principle class for visualization performance, first we instantiate the category and on this instance, the graph is directed. The barnes_hut methodology is then known as, which is a gravity mannequin. The from_nx methodology takes an current NetworkX graph as an argument and interprets it to PyVis, which is named in place.

from pyvis.community import Community

def convert_graph(graph):
    """
    Convert networkx to pyvis graph.
    """
    web = Community(
        peak="700px",
        width="100%",
        bgcolor="#222222",
        font_color="white",
        directed=True,
    )
    web.barnes_hut()
    web.from_nx(graph)
    return web

To render the Graph to the UI, we first create a singular consumer ID as we use the PyVis save_graph methodology to save lots of the HTML file for the graph on the server. The uuid ensures a singular file title, which is then learn into the streamlit UI and after the file is deleted.

import uuid
import contextlib
import os
import streamlit as st

def render_graphs(web):
    """
    Helper to render graph visualization from pyvis graph.
    """
    uuid4 = uuid.uuid4()
    file_name = f"./output/{uuid4}.html"
    with contextlib.suppress(FileNotFoundError):
        os.take away(file_name)
    web.save_graph(file_name)
    with open(file_name, "r", encoding="utf-8") as html_file:
        source_code = html_file.learn()
    st.elements.v1.html(source_code, peak=650, width=650)
    os.take away(file_name)

6. Graph RAG Utilizing LlamaIndex

Via graph retrieval-augmented technology, we will question our graph information straight, an instance may be seen within the prior screenshot. Extracted entities from the consumer question are regarded up within the graph to provide particular context to the AI to floor its response, as this info would probably not have been within the coaching corpus, and therefore any reply given would have had an elevated probability of being a hallucination.

We create a chat engine to cross a consumer’s earlier question historical past into the mannequin. Often, the Open AI API secret’s learn as an setting variable inside LlamaIndex — nevertheless, since that is user-submitted for our utility and we don’t wish to save customers’ Open AI credentials, we have to cross credentials to the LLM and embedding mannequin courses as key phrase arguments.

We then create an empty LlamaIndex Information Graph Index and populate the information graph by inserting triples. The triples come from traversing the sides of our NetworkX graph and calling the upsert_triplet_and_node methodology, which can create the triple and node in the event that they don’t exist already.

Because the graph is directed, we will interchange the themes and objects in order that the graph is traversable in both route. The chat engine makes use of the tree_summarize possibility for the response builder.

“Tree summarize response builder. This response builder recursively merges textual content chunks and summarizes them in a bottom-up trend (i.e. constructing a tree from leaves to root). Extra concretely, at every recursively step: 1. we repack the textual content chunks so that every chunk fills the context window of the LLM 2. if there is just one chunk, we give the ultimate response 3. in any other case, we summarize every chunk and recursively summarize the summaries.”— LlamaIndex Website

Calling the chat methodology with the consumer’s question and setting up the chat historical past from the Streamlit state object is included right here.

from llama_index.core import KnowledgeGraphIndex
from llama_index.core.schema import TextNode
from llama_index.embeddings.openai import OpenAIEmbedding
from llama_index.llms.openai import OpenAI
from llama_index.core.llms import ChatMessage, MessageRole
import streamlit as st
import utils.ui_utils as ui_utils  # pylint: disable=consider-using-from-import, import-error

def init_llama_index_graph(graph_nx, open_ai_api_key):
    """
    Assemble a information graph utilizing llama index.
    """
    llm = OpenAI(mannequin="gpt-3.5-turbo", api_key=open_ai_api_key)
    embed_model = OpenAIEmbedding(api_key=open_ai_api_key)

    graph = KnowledgeGraphIndex(
        [], llm=llm, embed_model=embed_model, api_key=open_ai_api_key
    )

    for subject_entity, object_entity in graph_nx.edges():
        predicate = graph_nx[subject_entity][object_entity].get("label", "pertains to")
        graph.upsert_triplet_and_node(
            (subject_entity, predicate, object_entity), TextNode(textual content=subject_entity)
        )
        graph.upsert_triplet_and_node(
            (object_entity, predicate, subject_entity), TextNode(textual content=subject_entity)
        )

    chat_engine = graph.as_chat_engine(
        include_text=True,
        response_mode="tree_summarize",
        embedding_mode="hybrid",
        similarity_top_k=5,
        verbose=True,
        llm=llm,
    )

    return chat_engine

def add_result_to_state(query, response):
    """
    Add mannequin output to state.
    """
    if response:
        graph_answers = st.session_state.get("graph_answers") or []
        graph_answers.append((query, response))
        st.session_state["graph_answers"] = graph_answers
    else:
        st.error("Question failed, please strive once more later.", icon="⚠️")

def query_llama_index_graph(query_engine, query):
    """
    Question llama index information graph utilizing graph RAG.
    """
    graph_answers = st.session_state.get("graph_answers", [])
    chat_history = []
    for question, reply in graph_answers:
        chat_history.append(ChatMessage(function=MessageRole.USER, content material=question))
        chat_history.append(
            ChatMessage(function=MessageRole.ASSISTANT, content material=reply)
        )

    if response := query_engine.chat(query, chat_history):
        add_result_to_state(query, response.response)

Equally, I initially explored a LangChain implementation, although throughout some experimentation, I made a decision to proceed wth the LlamaIndex-based method beforehand demonstrated. For reference, I’ve included this beneath whether it is helpful to you.

Within the curiosity of brevity, the reason is omitted, although it needs to be self-explanatory for the reader.

from langchain_community.chains.graph_qa.base import GraphQAChain
from langchain_community.graphs import NetworkxEntityGraph
from langchain_community.graphs.networkx_graph import KnowledgeTriple
from langchain_openai import ChatOpenAI
import streamlit as st

def add_result_to_state(query, response):
    """
    Add mannequin output to state.
    """
    if response:
        graph_answers = st.session_state.get("graph_answers") or []
        graph_answers.append((query, response))
        st.session_state["graph_answers"] = graph_answers
    else:
        st.error("Question failed, please strive once more later.", icon="⚠️")

def construct_graph_langchain(graph_nx, open_ai_api_key, query):
    """
    Assemble a information graph in Langchain and preform graph RAG.
    """
    graph = NetworkxEntityGraph()
    for node in graph_nx:
        graph.add_node(node)

    for subject_entity, object_entity in graph_nx.edges():
        predicate = graph_nx[subject_entity][object_entity].get("label", "pertains to")
        graph.add_triple(KnowledgeTriple(subject_entity, predicate, object_entity))

    llm = ChatOpenAI(
        api_key=open_ai_api_key, mannequin="gpt-4", temperature=0, max_retries=2
    )

    chain = GraphQAChain.from_llm(llm=llm, graph=graph, verbose=True)

    if response := chain.invoke({"question": query}):
        reply = response.get("end result")
        add_result_to_state(query, reply)

7. Linting With Pylint

Since I’ve left some feedback within the code to disable the linter within the examples above (examples are referenced from the GitHub repo), I assumed I’d cowl the subject of linting briefly.

For these unfamiliar, linting helps to verify your code for potential bugs and stylistic points. Linters routinely implement coding requirements.

To get began, set up Pylint by working the command.

pip set up pylint

Secondly, we have to create a .pylintrc file on the root of the challenge (we will additionally set default international and user-specific settings relying on the place we create the .pylintrc file). To do that, you will want to run.

pylint --generate-rcfile > .pylintrc

We are able to configure this file to suit our preferences by updating the default values throughout the .pylintrc file.

To run the linter manually, you should use.

pylint ./important.py && pylint ./**/*.py

When the Docker picture is constructed, it can routinely run Pylint and lift an error ought to it detect a difficulty with the code. This may be seen within the Dockerfile.

FROM python:3.10.16 AS base 

WORKDIR /app

COPY necessities.txt .

RUN pip set up --upgrade pip

RUN pip set up -r necessities.txt 

COPY . .

RUN mkdir -p /app/output

RUN pylint ./important.py && pylint ./**/*.py

RUN python -m unittest -v checks.test_ukri_utils.Testing

CMD ["streamlit", "run", "./main.py"]

A preferred formatter that you may additionally discover helpful is Black —

“Black is a PEP 8 compliant opinionated formatter. Black reformats complete recordsdata in place.”

Working Black will routinely resolve a number of the points that will be raised by the linter.

8. Streamlit Group Cloud Demo App

With Streamlit Community Cloud, anybody can host their utility at no cost. You probably have an utility you’d prefer to deploy, you may observe this tutorial.

To see the hosted demo, please click on the hyperlink beneath.

https://governmentfundinggraph.streamlit.app

Thanks for studying my article — as promised, you could find all of the code within the GitHub repo here.

Any and all suggestions is efficacious to me because it gives route for my future tasks. For those who discovered this text helpful, please let me know.

You too can discover me over on LinkedIn in case you have particular questions.

Interested by open-source AI grant writing tasks? Join our mailing record here.

*All photos, except in any other case famous, are by the creator.

References