Context formats

Every surface (log(), @trace, the adapter) accepts the context argument in three shapes, so it fits whatever your pipeline already holds. All three normalize to the same internal ContextChunk list.

1. List of strings

The simplest form — just the chunk texts, in retrieval order:

context = [
    "Divide 72 by the annual rate to estimate doubling time.",
    "At 8%, money doubles in roughly 9 years.",
]

Each string becomes a chunk with rank set to its position (0 = top).

2. List of dicts

Use dicts when you have ranking, source, or score metadata:

context = [
    {"text": "Divide 72 by the annual rate…", "rank": 0, "source": "finance_101.md", "score": 0.91},
    {"text": "At 8%, money doubles in ~9 years.", "rank": 1, "source": "finance_101.md", "score": 0.87},
]

Note: the field is text. If your chunks are LangChain Document objects (which use page_content), map them first — e.g. [d.page_content for d in docs] — or use the LangChain adapter, which reads page_content for you.

3. List of ContextChunk

The typed form, if you're already constructing them:

from veralith import ContextChunk
 
context = [
    ContextChunk(text="Divide 72 by the annual rate…", rank=0, source="finance_101.md", score=0.91),
]

ContextChunk fields: text: str (required), rank: int (0 = top), source: str | None, score: float | None.

Why source and score help

They're optional, but sending them makes diagnoses richer: the dashboard can show which document a fabricated claim ignored, and low-scoring chunks that still got used are a common retrieval smell.