Extracting Entities by Hand

Where we left off

Lessons 1 and 2 established the problem: a multi-hop question needs facts connected across documents, and similarity search can't find that connection on its own. Building a knowledge graph starts with the same first step humans use when reading a document and mentally noting "who and what is this about": pulling out the entities, the people, places, and things a document talks about.

Why prompt Gemini for this instead of writing a parser

You might assume this needs proper NLP tooling, a named-entity-recognition model, part-of-speech tagging, something specialized. It doesn't, not at this scale. A general-purpose language model like Gemini is already extremely good at "read this paragraph, tell me who and what it's about," because that's a language-understanding task, not a specialized statistical one. Asking a general model to do this well just requires being specific about the shape of the answer you want back, which is what the prompt below does.

The code, piece by piece

prompt = f"""List the named entities (people, places, and things - not
generic nouns) mentioned in the text below. Return them as a JSON array
of short strings, using each entity's most natural full name...

Text:
{text}"""

response = client.models.generate_content(
    model=CHAT_MODEL,
    contents=prompt,
    config=types.GenerateContentConfig(response_mime_type="application/json"),
)
entities = json.loads(response.text)

Two things matter here. First, the prompt explicitly excludes "generic nouns": without that instruction, a model will happily list "sensor," "garage," and "readings" as entities, which are real words but not the kind of thing you'd draw an edge to or from in a graph, they're categories, not identifiable things. Second, response_mime_type="application/json" isn't a formatting nicety, it's what makes this reliable: instead of hoping Gemini's free-text response happens to be parseable, this setting constrains the model to emit valid JSON, which json.loads can then turn straight into a Python list, no regex or string-splitting involved.