llm-convrec

KO | EN

gitlite — search

#python #llm #android #machine-learning #linux #java #nodejs #windows #cli #vue #awesome-list #mcp #openai #ios

(★ 63)

LLM-based Conversational Recommendation Architecture

llm-convrec Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# llm-convrec

[![Streamlit Website](https://img.shields.io/badge/Streamlit-Demo-FF4B4B?logo=streamlit&logoColor=white)](https://ra-rec.streamlit.app)
[![Colab Live Demo](https://img.shields.io/badge/Colab-Demo-F9AB00?logo=googlecolab&logoColor=white)](https://colab.research.google.com/drive/1oboNxF_XpSpa3MbTiVukObmFHP6l0bzD?usp=sharing)
[![ACM SIGIR Digital Library](https://img.shields.io/badge/ACM%20SIGIR-0085CA?logo=acm&logoColor=fff&style=flat)](https://dl.acm.org/doi/abs/10.1145/3626772.3657670)
[![License: MIT](https://img.shields.io/badge/License-MIT-success.svg)](https://github.com/D3Mlab/llm-convrec/tree/main?tab=MIT-1-ov-file#readme)

Conversational Flow:  
<img width="800" alt="Screenshot 2025-03-26 at 1 16 56 PM" src="https://github.com/user-attachments/assets/e894e6d3-284e-468a-a442-4c269805076a" />

</div>

## Introduction: A Semi-Structured Conversational Recommendation System

LLM-ConvRec is a **semi-structured** conversational recommendation system built on **retrieval-augmented generation** (RAG). It combines the **flexibility of LLM-driven** dialogue with the **reliability of grounded retrieval**, enabling high-quality, natural conversations backed by a factual item database and user reviews.

Unlike end-to-end black-box models, LLM-ConvRec maintains an **explicit internal state** that tracks user preferences and constraints. This structure improves response consistency, memory retention, and control—mitigating common failure modes such as hallucination, repetition, or loss of context.

The system operates in four stages per turn: **intent classification, state update, action selection**, and **LLM-based response generation**, with LLM prompting informed by both retrieved knowledge and structured memory. The result is a system that:

- Produces fluent, adaptive responses via LLMs.

- Grounds recommendations in a **real database** of items and reviews.

- Maintains **interpretable and editable conversation state** for traceability and personalization.

- Enforces **structural constraints** to ensure safe, relevant, and goal-driven dialogue.

This architecture allows LLM-ConvRec to scale across domains with minimal reconfiguration while retaining domain-specific customization and control.

<div align="center">
  
## Example Conversation
<img width="522" alt="Screenshot 2025-03-26 at 1 41 42 PM" src="https://github.com/user-attachments/assets/a817104a-230d-4e05-aa1a-df46677e0660" />

</div>

## Table of Content

- [Introduction: A Semi-Structured Conversational Recommendation System](#introduction-a-semi-structured-conversational-recommendation-system)
- [Example Conversation](#example-conversation)
- [Quick Start](#quick-start)
- [Installation and Running the System](#installation-and-running-the-system)
  - [1. Clone the GitHub Repository](#1-clone-the-github-repository)
  - [2. Navigate to the Project Directory](#2-navigate-to-the-project-directory)
  - [3. Install the Required Packages](#3-install-the-required-packages)
  - [4. Obtaining an OpenAI API Key and Configuring the .env file](#4-obtaining-an-openai-api-key-and-configuring-the-env-file)
  - [5. Run the System](#5-run-the-system)
- [Overall Conversation Flow](#overall-conversation-flow)
  - [1. Intent Classification](#1-intent-classification)
  - [2. State Update](#2-state-update)
  - [3. Action Classification](#3-action-classification)
  - [4. Response Generation](#4-response-generation)
- [Domain Initialization and Customization](#domain-initialization-and-customization)
  - [Quick Start](#quick-start)
  - [1. Constraints Configuration](#1-constraints-configuration)
  - [2. Few shots for prompts](#2-few-shots-for-prompts)
  - [3. Hard-Coded Responses](#3-hard-coded-responses)
  - [4. Filter Configs](#4-filter-configs)
  - [5. Data](#5-data)
  - [6. User Defined Classes](#6-user-defined-classes)
  - [7. Domain Specific Config](#7-domain-specific-config)

## Quick Start

Here is the link to the Google Colab for a quick start:

https://apoj.short.gy/d3m-llm-convrec-demo

## Installation and Running the System

Before you can use the system, you must first ensure that you have Python (version 3.7 or higher) installed on your machine. You will also need pip for installing the necessary Python packages.

### 1. Clone the GitHub Repository

Clone the repository from GitHub to your local machine by running the following command in your terminal:

**git clone https://github.com/D3Mlab/llm-convrec.git**

### 2. Navigate to the Project Directory

Once you've cloned the repository, use the command line to navigate into the project's directory:

**cd llm-convrec**

### 3. Install the Required Packages

The project has a number of dependencies that need to be installed. These are listed in the `requirements.txt` file. To install these dependencies, run the following command in your terminal:

**pip install -r requirements.txt**

### 4 Obtaining an OpenAI API Key and Configuring the .env file

Before you can run the system, you need an API key from OpenAI. This key enables the model to interact with the OpenAI's servers to process and generate conversational responses.

Here are the steps to obtain an API key:

1. Visit OpenAI's website and create an account or log in if you already have an account.

2. Navigate to the 'API Keys' section in your account settings.

3. Click on the 'Create a new API key' button.

4. Name your key and click on 'Create secret key' to generate your new API key.

**Note that the system cannot be used without entering your credit card information to your OpenAI account since the system interactions exceed the free API limitations. A typical conversation costs < $0.02.**

After you have the API key, you need to configure your `.env` file:

1. Create a new file in your project root directory and name it `.env`.

2. Inside the `.env` file, create a new line and write `OPENAI_API_KEY=`, and then paste your API key after the equals sign. For example:

OPENAI_API_KEY='sk1234567890abcdef`.

4. Save the `.env` file.

Please ensure you do not upload your `.env` file to public repositories to keep your OpenAI API key secure.

### 5. Run the System

If you want to run the restaurant demo execute following command in the terminal:

```
python restaurant_main.py
```
The restaurant demo uses geocoding in order to verify location. It currently uses `Nominatim`, but can be configured to use Google Maps API. To do this, you must create a Google API Key. To get the API Key, go to https://console.developers.google.com/ and navigate to 'Credentials' tab after logging in to your google cloud account and then click on 'CREATE CREDENTIALS'. Make sure to have both Geocoding API and Time Zone API services enabled for this API key.

Inside the `.env ` file, create a new line and write `GOOGLE_API_KEY=` and then paste the API key in after the equals sign. For example:
GOOGLE_API_KEY = hghrjkdkxhgyrujjedksdk

Or, here is the link to the Google Colab for a quick start:

https://apoj.short.gy/d3m-llm-convrec-demo

## Overall Conversation Flow

The LLM-ConvRec system follows a precise process during each conversation turn to ensure the highest quality of responses. Let's break down this process, from the initial user input to the final recommender response:

Examples to demonstrate these categories are from the restaurant recommendation domain.

### 1. Intent Classification

The conversation begins with user input. This input is passed to an Intent Classifier, which determines the user's intent. The intent could be any or multiple of the following:

- **Provide Preference:** The user expresses a preference or interest. (Example: "I'd like to eat some sushi.")
- **Inquire:** The user asks a question about a specific item or detail. (Example: "What's on their menu? Does the restaurant have a patio?")
- **Accept/Reject Recommendation:** The user responds to a previous recommendation made by the system.(Example: "Sure, that first one sounds good!")

The prompts used for intent classification can be found here: https://github.com/D3Mlab/llm-convrec/tree/main/prompt_files/user_intent_prompts

### 2. State Update

After identifying the user's intent, the system updates its internal state. This state stores critical information gathered during the conversation, including:

- **User's Preferences and Constraints:** These include their location, budget, dietary restrictions, etc.
- **Current Item of Interest:** The specific item (e.g., restaurant) the user is currently referring to or interested in.
- **Accepted and Rejected Items:** The system keeps track of the items that the user has accepted or rejected.

The prompts used for preferences and constraints extraction can be found here: https://github.com/D3Mlab/llm-convrec/tree/main/prompt_files/constraints_prompts

The prompts used for current, accepted, and rejected item extraction can be found here: https://github.com/D3Mlab/llm-convrec/tree/main/prompt_files/items_extractor_prompts

### 3. Action Classification

With the updated state, the system then decides the action to take next. This could be any of the following:

- **Request Information:** The system may ask the user for additional details to refine its understanding or recommendations. (Example: "Can you provide your location?")
- **Give Recommendation:** The system might suggest an item that matches the user's stated preferences.
- **Answer a Question:** If the user asked a question in their last utterance, the system would provide an appropriate answer.

### 4. Response Generation

Following the action classification, the system produces a response that corresponds with the determined action. This response creation leverages retrieval augmented information retrieval, which is particularly employed in 'recommend' and 'answer' actions.

The response is not only contextually in sync with the ongoing dialogue, but also respects the semi-structured conversation pattern that the system adheres to.

The prompts used for recommendation and explanation can be found here: https://github.com/D3Mlab/llm-convrec/tree/main/prompt_files/recaction_prompts/recommend_prompts

The prompts used for answering questions can be found here: https://github.com/D3Mlab/llm-convrec/tree/main/prompt_files/recaction_prompts/answer_prompts

## Domain Initialization and Customization

This system is designed to be flexible and adaptable, allowing you to initialize and customize your own domain. With a configuration process involving providing some key files, you can utilize our robust system architecture tailored to your specific needs.

## Quick Start

If you're looking to get started quickly, we've already set up a pre-configured domain: **Restaurant**.

- **Restaurant Domain:** This domain utilizes a Yelp dataset containing all Edmonton restaurants. The domain is already initialized and ready to use, providing a wide range of restaurant data.

Due to data available, the location is restricted to Edmonton, Canada.

Here is the link to the Google Colab for a quick start:

https://apoj.short.gy/d3m-llm-convrec-demo

## Setting Up Customized Domain

The domain specific files reside in the domain_specific folder, where there are two subfolders: `classes` and `configs`.

1. The `classes` folder stores all of the user defined domain specific classes (more information on this below), which are completely optional but allow for a better user experience.
2. The `config` folder stores all of the domain specific configs necessary in order to make a recommendation.

If you want to follow our structure, create a folder representing the new domain in both the `config` and `classes` folders. For example, we named the restaurant domain folders `restaurant_configs` (`domain_specific/configs/restaurant_configs`) and restaurants (`domain_specific/classes/restaurants`).

Note that key files should be stored in domain_specific folder, with all CSV and yaml files being stored in `domain_specific/configs/<YOUR_DOMAIN_CONFIG_FOLDER>`.

## 1. Constraints Configuration

To provide personalized recommendations, the LLM-ConvRec system takes into account user constraints that can be both explicit (provided directly by the user) or implicit (derived from the user's input). For efficient constraint management, it is crucial to set up a `constraints_config.csv` that defines the various constraints and their properties.

The `constraints_config.csv` file should include the following columns:

- **key**: The constraint's key name.
- **description**: Description of the constraint.
- **is_cumulative**: A Boolean value (TRUE or FALSE) indicating if the constraint's value is cumulative. If `is_cumulative` is TRUE, the system appends the newly extracted values to the existing values of the constraint rather than overwriting them. If FALSE, any newly identified value replaces the previous value.
- **in_explanation**: A Boolean value (TRUE or FALSE) indicating if the constraint's value should be considered when giving explanation about the item to the user, during recommendation stage.
- **default_value**: The default value of the constraint when it is not specified by the user.

Below is an example of how the `constraints_config.csv` file should look:

| key          | description                                                                                                                                                                                                     | is_cumulative | in_explanation | default_value |
| ------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------- | -------------- | ------------- |
| location     | The desired location of the restaurants.                                                                                                                                                                        | FALSE         | FALSE          | None          |
| cuisine type | The desired specific style of cooking or cuisine offered by the restaurants (e.g., "Italian", "Mexican", "Chinese"). This can be implicitly provided through dish type (e.g "italian" if dish type is "pizza"). | FALSE         | TRUE           | None          |
| dish type    | The desired menu item or dish in the restaurant that user shows interests.                                                                                                                                      | TRUE          | TRUE           | None          |
| type of meal | The desired category of food consumption associated with specific times of day (e.g., "breakfast", "lunch", "dinner").                                                                                          | TRUE          | TRUE           | None          |

The configuration of these constraints will allow the system to capture user preferences more accurately, leading to more personalized and relevant recommendations.

## 2. Few shots for prompts

Few-shot examples are crucial in training the LLM-ConvRec system. They are a set of input-output pairs that demonstrate the type of behavior we want the system to exhibit. In the context of our system, few-shot examples help train the classifiers and provide the necessary prompts for information extraction.

When providing few-shot examples, make sure that they are representative of the tasks you want the model to perform. For instance, if you want the system to recognize when a user is expressing a preference, include examples where users express preferences in different ways.

Please note that these examples are illustrative. The content of your files will be determined by the specific nature of your domain and the complexity of the user's queries.
Remember, the quality of the few-shot examples can significantly impact the performance of the system. Carefully curating these examples will lead to a more responsive and accurate conversational system.

### 2.1 Few-shots for Intent Classification Prompts

For effective intent classification, few-shot examples must be provided for each intent. This should be done in the form of CSV files with two columns: `User Input` and `Response`. `User Input` should contain examples of user utterances, while `Response` indicates whether the input corresponds to the respective intent (True) or not (False).

#### 2.1.1 `accept_classification_fewshots.csv`:

This file should contain examples of user utterances that express acceptance of a recommendation.

| User Input                       | Response |
| -------------------------------- | -------- |
| That sounds good, let's go there | True     |
| I don't like that type of food   | False    |

#### 2.1.2 `reject_classification_fewshots.csv`:

This file should contain examples where the user rejects a recommendation.

| User Input                           | Response |
| ------------------------------------ | -------- |
| No, I don't want to go to that place | True     |
| Sure, that sounds nice               | False    |

#### 2.1.3 `inquire_classification_fewshots.csv`:

This file should contain examples where the user is inquiring or asking a question.

| User Input                           | Response |
| ------------------------------------ | -------- |
| What's on their menu?                | True     |
| I think we should try something else | False    |

### 2.2 Few-shots for Constraint Updater prompt

The 'constraints_updater_fewshots.csv' file plays an instrumental role in enhancing the personalization of recommendations by keeping track of explicit and implicit user preferences. It's fundamental in identifying and storing the constraints from the user's conversation, categorizing them as hard or soft constraints. Hard constraints are strict requirements, such as a specific location or cuisine type, which must be adhered to. On the other hand, soft constraints denote user preferences that are desirable but not necessarily mandatory, like a preference for a place offering free parking or a patio.

#### 2.2.1 `constraints_updater_fewshots.csv`:

This file should be structured with five columns: 'user_input', 'old_hard_constraints', 'old_soft_constraints', 'new_hard_constraints', and 'new_soft_constraints'.

- 'user_input' represents a user's utterance.
- 'old_hard_constraints' and 'old_soft_constraints' denote the previously identified hard and soft constraints, respectively.
- 'new_hard_constraints' and 'new_soft_constraints' represent the updated set of hard and soft constraints after processing the user's input.

Here is an example of the structure of this CSV file:

| user_input                                 | old_hard_constraints                                              | old_soft_constraints                                | new_hard_constraints                                                           | new_soft_constraints                                         |
| ------------------------------------------ | ----------------------------------------------------------------- | --------------------------------------------------- | ------------------------------------------------------------------------------ | ------------------------------------------------------------ |
| pizza and pasta                            | "location=["Toronto"]"                                            |                                                     | "location=["toronto"], cuisine type=["italian"], dish type=["pizza", "pasta"]" |                                                              |
| does it have a patio?                      | "location=["jasper avenue, edmonton"], cuisine type=["japanese"]" | "price range=["moderate"], others=["free parking"]" | "location=["jasper avenue, edmonton"], cuisine type=["japanese"]"              | "price range=["moderate"], others=["free parking", "patio"]" |
| What kind of menu does I Love Sushi offer? | "location=["jasper avenue"], cuisine type=["japanese"]"           |                                                     | "location=["jasper avenue"], cuisine type=["japanese"]"                        |

The ability to track and update these evolving constraints allows the system to fine-tune its recommendations, significantly enhancing the overall conversation experience.

### 2.4 Few-shots for 'Answer' Recommender Action Prompts

This section provides details about the few-shot prompt CSV files required for the 'Answer' recommender action.

#### 2.4.1 `answer_extract_category_fewshots.csv`

This file helps in mapping user queries to metadata categories. It needs two columns: `input` (user's question) and `output` (one of the metadata field or one of the key in optional field).

| input                                        | output          |
| -------------------------------------------- | --------------- |
| What's their addresses?                      | address         |
| Can you recommend any dishes or specialties? | none            |
| Can I make a reservation?                    | HasReservations |
| What are the meals it's known for?           | PopularMeals    |

#### 2.4.2 `answer_ir_fewshots.csv`

This file instructs the model to extract answers from reviews based on the user's question. It requires `question` (user's question), `information` (reviews retrieved by the system), and `answer` (the answer to the question derived from the provided information).

Note that the answer "I do not know." to the corresponding prompt is a special one that triggers a default response that is responsible for the situation where the question is not relevant to the information we have access to. Hence, one of the fewshot examples should always be an example question that should be answered with "I do not know.". This is illustrated in the below example too.

| question                                | information                                     | answer         |
| --------------------------------------- | ----------------------------------------------- | -------------- |
| Do they have a slide in the restaurant? | I really like this place. They have great food. | I do not know. |

#### 2.4.3 `answer_separate_questions_fewshots.csv`

This file aids the system in breaking down complex user queries into simpler, individual questions. It requires `question` (user's question) and `individual_questions` (decomposed questions).

| question                                                       | individual_questions                                                                                  |
| -------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------- |
| Do they have wine?                                             | Do they have wine?                                                                                    |
| What are dishes, cocktails and types of wine do you recommend? | What dishes do you recommend?\nWhat cocktails do you recommend?\nWhat types of wine do you recommend? |

## 3. Hard-Coded Responses

The `hard_coded_responses.csv` file includes specific, structured responses that the system should use under certain conditions. This file has three columns: `Action`, `Response`, and `Constraints`.

- **Action**: This column refers to the action that the system should take.
- **Response**: This column provides the exact response that should be given when the action is chosen.
- **Constraints**: This column specifies one or more constraints. If any of these constraints are missing in the system's state, then the corresponding response will be prioritized. If no constraints are applicable, this field should be left empty.

Descriptions of each action:

- **PostAcceptanceAction**: The response given after the user accepts a recommendation.
- **PostRejectionAction**: The response given after the user rejects a recommendation.
- **RequestInformation**: The system is requesting additional information from the user. The required information type depends on the 'Constraints' column.
- **DefaultResponse**: Used when the system failed to classify an intent and cannot decide on any specific action until the user provides more information.
- **NoRecommendation**: This response is given when the system cannot find a restaurant that meets the user's constraints.
- **NoAnswer**: Used when the system cannot answer the user's question with the information it has retrieved.
- **InitMessage**: This message is given to the user at the beginning.

Here are some examples:

| Action               | Response                                                                                                                                                                     | Constraints             |
| -------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------- |
| PostAcceptanceAction | Great! If you need any more assistance, feel free to ask.                                                                                                                    |                         |
| PostRejectionAction  | I'm sorry that you did not like the recommendation. Is there anything else I can assist you with?                                                                            |                         |
| RequestInformation   | Could you provide the location?                                                                                                                                              | location                |
| RequestInformation   | Could you provide the cuisine type or dish type?                                                                                                                             | cuisine type, dish type |
| RequestInformation   | Do you have any other preferences?                                                                                                                                           |                         |
| DefaultResponse      | Could you provide more information?                                                                                                                                          |                         |
| NoRecommendation     | Sorry, there is no restaurant that matches your constraints.                                                                                                                 |                         |
| NoAnswer             | Please only ask questions about previously recommended restaurant.                                                                                                           |                         |
| InitMessage          | Hello there! I am a restaurant recommender. Please provide me with some preferences for what you are looking for. For example, location, cuisine type, or dish type. Thanks! |                         |

By providing these hard-coded responses, you can control the behavior of the system and ensure that the conversation flow remains on track.

## 4. Filter Configs

### 4.1 `filter_config.csv`

The `filter_config.csv` is a configuration file that allows you to specify the filters you want to apply in the system. This CSV file consists of three columns: `type_of_filter`, `key_in_state` and `metadata_field`.

- `type_of_filter`: This column specifies the type of filter you want to use. Valid types include `"exact word matching"`, `"word in"`, `"value range"`, and `"item"`.

- `key_in_state`: This column lists the keys in the hard constraint that you want to check against `metadata_field`. Depending on the filter type, this could be a list of keys or a single key.

- `metadata_field`: This column contains the field you want to check against `key_in_state`.

Here is an example of what this file might look like:

| type_of_filter | key_in_state            | metadata_field |
| -------------- | ----------------------- | -------------- |
| word in        | cuisine type, dish type | categories     |
| item           | recommended_items       | name           |

Below, we describe several types of filters that can be used in the system. They allow you to filter out items based on various criteria.

### 1.Exact Word Matching Filter

This filter retains an item if any value in `key_in_state` exactly matches a value in `metadata_field`. This filter is case insensitive.

**Type of filter:** `"exact word matching"`

**key_in_state:** List of keys in the hard constraint you want to check against `metadata_field`.

**metadata_field:** The field you want to check against `key_in_state`. The value in this field should be a string, a list-like string (e.g. “A, B, C”), or a list. If it is neither a string nor a list, the item will be retained.

### 2.Word In Filter

This filter retains an item if any value in `key_in_state` is present in `metadata_field`, or vice versa. It also considers plural forms. This filter is case insensitive.

**Type of filter:** `"word in"`

**key_in_state:** List of keys in the hard constraint you want to check against `metadata_field`.

### 3.Value Range Filter

This filter retains an item if any value in `metadata_field` falls within a value range in `key_in_state`, or if any value range in `metadata_field` overlaps with a value range in `key_in_state`.

**Type of filter:** `"value range"`

**key_in_state:** A key in the hard constraint containing the value ranges to check against `metadata_field`. Value ranges should be in the format of "(lower bound value (with units))-(upper bound value (with units))".

**metadata_field:** The field you want to check against `key_in_state`. The value in this field should be a string (either a value or a value range), a list-like string of values (e.g. “A, B, C”), or a list of values. If it is neither a string nor a list, the item will be retained.

### 4. Item Filter

This filter retains an item if it is not in the list of recommended items specified by `key_in_state`.

**Type of filter:** `"item"`

**key_in_state:** A key in the state manager that contains a list of recommended items to check against `metadata_field`.

**metadata_field:** The field to use for checking whether an item is not in the value of `key_in_state`. Must be either `item_id` or `name`.

## 5. Data

The LLM-ConvRec system requires two main types of data: metadata and reviews.

### 5.1 Metadata

The metadata must include name (name) and unique item identifiers (item_id) as keys(columns). Each item can have various other keys representing different types of metadata, such as location, type of cuisine, cost, etc. It is not necessary for all items to have a value for every metadata field.

In addition, the metadata must have an "optional" key where the value contains the key-value pairs of optional categories.

An example of a metadata structure is as follows:

{
    "item_id": "001",
    "name": "Brits Fish & Chips",
    "address": "6940 77 Street NW",
    "city": "Edmonton",
    "categories": ["Fish & Chips", "Restaurants"],
    "optional": {"GoodForKids": "True", "OutdoorSeating": "False"}
    }...

### 5.2 Review data

The review data must have two keys(columns): "item_id" and "Review".

An example of a review file structure is as follows:

| item_id | Review             |
| ------- | ------------------ |
| 001     | great! ...         |
| 001     | I like it ...      |
| 002     | this restaurant... |

Note that the order of item_id in the review data must correspond to the order in the metadata.

## 6. User defined classes

User defined classes are used to implement domain specific tasks, for example, merging constraints in a specialized way.
You can create your own implementations of the following classes:

### Constraint Merger:

Defines special way to merge constraint

Example implementation:  `Location Merger` (merges two location using geocoding)

Path to abstract class: `state/constraints/constraint_merger.py`

### Status:

Allows constraint to have custom status and corresponding to response from the recommender

Example implementation: `Location Status` (location can be “invalid”, “valid”, or “specific”)

Path to abstract class: `state/status.py`

### Filter:

Defines a way of filtering item based on the constraints and metadata

Path to abstract class: `information_retriever/filter/filter.py`

Example implementation: `Location Filter` (filters based off of location)

## 7. Domain Specific Config

Once all the mentioned csv files and data files have been created, they can be put into one folder under the `domain_specific/configs` folder, and then modify the `PATH_TO_DOMAIN_CONFIGS` to be:

```
PATH_TO_DOMAIN_CONFIGS: "domain_specific/configs/<YOUR_DOMAIN_CONFIG_FOLDER>"
```

Then, create a `domain_specific_config.yaml` file and modify the file paths, similar to the below example:

```
DOMAIN: < DOMAIN_NAME_AS_A_NOUN >

EXPLANATION_METADATA_BLACKLIST: < all metadata keys that should be ignored when giving explanation of the item to the user, during recommendation stage >

INQUIRE_CLASSIFICATION_FEWSHOTS_FILE: "inquire_classification_fewshots.csv"

ACCEPT_CLASSIFICATION_FEWSHOTS_FILE: "accept_classification_fewshots.csv"

REJECT_CLASSIFICATION_FEWSHOTS_FILE: "reject_classification_fewshots.csv"

CONSTRAINTS_CATEGORIES: "constraints_config.csv"

CONSTRAINTS_UPDATER_FEWSHOTS: "constraints_updater_fewshots.csv"

ACCEPTED_ITEMS_EXTRACTOR_FEWSHOTS_FILE: "accepted_items_extractor_fewshots.csv"

REJECTED_ITEMS_EXTRACTOR_FEWSHOTS_FILE: "rejected_items_extractor_fewshots.csv"

CURRENT_ITEMS_EXTRACTOR_FEWSHOTS_FILE: "current_items_extractor_fewshots.csv"

ANSWER_EXTRACT_CATEGORY_FEWSHOTS_FILE: "answer_extract_category_fewshots.csv"

ANSWER_IR_FEWSHOTS_FILE: "answer_ir_fewshots.csv"

ANSWER_SEPARATE_QUESTIONS_FEWSHOTS_FILE: "answer_separate_questions_fewshots.csv"

HARD_CODED_RESPONSES_FILE: "hard_coded_responses.csv"

FILTER_CONFIG_FILE: "filter_config.csv"

PATH_TO_ITEM_METADATA: "data/item_metadata.json"

PATH_TO_REVIEWS: "data/items_reviews.csv"

PATH_TO_EMBEDDING_MATRIX: "data/reviews_embedding_matrix.pt"

PATH_TO_DATABASE: "data/database.faiss"
```
If you do not want to follow our structure, then modify the paths above to where you stored the corresponding files.

llm-convrec

llm-convrec Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

llm-convrec

llm-convrec Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README