Cypher Evaluation Chain

return RunnableSequence.from<
  CypherEvaluationChainInput,
  CypherEvaluationChainOutput
>([
  RunnablePassthrough.assign({
    // Convert errors into an LLM-friendly list
    errors: ({ errors }) => {
      if (
        errors === undefined ||
        (Array.isArray(errors) && errors.length === 0)
      ) {
        return "";
      }

      return `Errors: * ${
        Array.isArray(errors) ? errors?.join("\n* ") : errors
      }`;
    },
  }),
  // ...
]);

return RunnableSequence.from<
  CypherEvaluationChainInput,
  CypherEvaluationChainOutput
>([
  RunnablePassthrough.assign({
    // Convert errors into an LLM-friendly list
    errors: ({ errors }) => {
      if (
        errors === undefined ||
        (Array.isArray(errors) && errors.length === 0)
      ) {
        return "";
      }

      return `Errors: * ${
        Array.isArray(errors) ? errors?.join("\n* ") : errors
      }`;
    },
  }),
  prompt,
  llm,
  new JsonOutputParser<CypherEvaluationChainOutput>(),
]);

export default async function initCypherEvaluationChain(
  llm: BaseLanguageModel
) {
  // Prompt template
  const prompt = PromptTemplate.fromTemplate(`
    You are an expert Neo4j Developer evaluating a Cypher statement written by an AI.

    Check that the cypher statement provided below against the database schema to check that
    the statement will answer the user's question.
    Fix any errors where possible.

    The query must:
    * Only use the nodes, relationships and properties mentioned in the schema.
    * Assign a variable to nodes or relationships when intending to access their properties.
    * Use \`IS NOT NULL\` to check for property existence.
    * Use the \`elementId()\` function to return the unique identifier for a node or relationship as \`_id\`.
    * For movies, use the tmdbId property to return a source URL.
      For example: \`'https://www.themoviedb.org/movie/'+ m.tmdbId AS source\`.
    * For movie titles that begin with "The", move "the" to the end.
      For example "The 39 Steps" becomes "39 Steps, The" or "the matrix" becomes "Matrix, The".
    * For the role a person played in a movie, use the role property on the ACTED_IN relationship.
    * Limit the maximum number of results to 10.
    * Respond with only a Cypher statement.  No preamble.

    Respond with a JSON object with "cypher" and "errors" keys.
      * "cypher" - the corrected cypher statement
      * "corrected" - a boolean
      * "errors" - A list of uncorrectable errors.  For example, if a label,
          relationship type or property does not exist in the schema.
          Provide a hint to the correct element where possible.

    Fixable Example #1:
    * cypher:
        MATCH (a:Actor {{name: 'Emil Eifrem'}})-[:ACTED_IN]->(m:Movie)
        RETURN a.name AS Actor, m.title AS Movie, m.tmdbId AS source,
        elementId(m) AS _id, m.released AS ReleaseDate, r.role AS Role LIMIT 10
    * errors: ["Variable \`r\` not defined (line 1, column 172 (offset: 171))"]
    * response:
        MATCH (a:Actor {{\name: 'Emil Eifrem'}})-[r:ACTED_IN]->(m:Movie)
        RETURN a.name AS Actor, m.title AS Movie, m.tmdbId AS source,
        elementId(m) AS _id, m.released AS ReleaseDate, r.role AS Role LIMIT 10


    Schema:
    {schema}

    Question:
    {question}

    Cypher Statement:
    {cypher}

    {errors}
  `);

  return RunnableSequence.from<
    CypherEvaluationChainInput,
    CypherEvaluationChainOutput
  >([
    RunnablePassthrough.assign({
      // Convert errors into an LLM-friendly list
      errors: ({ errors }) => {
        if (
          errors === undefined ||
          (Array.isArray(errors) && errors.length === 0)
        ) {
          return "";
        }

        return `Errors: * ${
          Array.isArray(errors) ? errors?.join("\n* ") : errors
        }`;
      },
    }),
    prompt,
    llm,
    new JsonOutputParser<CypherEvaluationChainOutput>(),
  ]);
}