# Context packs

A context pack is the text that your application sends to a large language model.

Biblicus keeps two things separate:

- Retrieval returns **evidence** as structured objects with provenance.
- Context pack building turns evidence into **context pack text** using an explicit policy.

This separation makes retrieval repeatable and testable, while keeping context formatting as an explicit surface you can change, compare, and evaluate.

## Minimal policy

The minimal policy is: join evidence text blocks with a separator.

In Python:

```python
from biblicus.context import ContextPackPolicy, build_context_pack

policy = ContextPackPolicy(join_with="\n\n")
context_pack = build_context_pack(result, policy=policy)
print(context_pack.text)
```

### Output structure

Context pack building returns a structured result you can inspect:

```json
{
  "text": "item_id: ...",
  "evidence_count": 2,
  "blocks": [
    {
      "evidence_item_id": "ITEM_ID",
      "text": "item_id: ITEM_ID\nsource_uri: ...",
      "metadata": {
        "item_id": "ITEM_ID",
        "source_uri": "file:///...",
        "score": 0.42,
        "stage": "retrieve"
      }
    }
  ]
}
```

`blocks` keeps a per-evidence record so you can debug how the final text was assembled.

### Before and after example

Given two evidence blocks, compare how different policies change the output:

```python
policy = ContextPackPolicy(join_with="\n\n", ordering="rank", include_metadata=False)
context_pack = build_context_pack(result, policy=policy)
print(context_pack.text)
```

With metadata enabled and score ordering:

```python
policy = ContextPackPolicy(join_with="\n\n", ordering="score", include_metadata=True)
context_pack = build_context_pack(result, policy=policy)
print(context_pack.text)
```

The first output keeps the original ranking and clean text blocks. The second output reorders by score and adds
explicit metadata lines for inspection.

## Policy surfaces

Context pack policies make ordering and formatting explicit.

### Ordering

Use `ordering` to control how evidence blocks are arranged before joining:

- `rank`: use the evidence rank as provided by retrieval.
- `score`: sort by score (descending) and then item identifier.
- `source`: group by source uniform resource identifier, then sort by score.

### Metadata inclusion

Set `include_metadata=True` to prepend metadata to each block. Metadata includes:

- `item_id`
- `source_uri`
- `score`
- `stage`

### Character budgets

Character budgets drop trailing blocks until the context pack fits the specified limit. This keeps context shaping
deterministic without relying on a tokenizer.

In Python:

```python
from biblicus.context import CharacterBudget, ContextPackPolicy, fit_context_pack_to_character_budget

policy = ContextPackPolicy(join_with="\n\n", ordering="score", include_metadata=True)
fitted = fit_context_pack_to_character_budget(context_pack, policy=policy, character_budget=CharacterBudget(max_characters=500))
print(fitted.text)
```

## Command-line interface

The command-line interface can build a context pack from a retrieval result by reading JavaScript Object Notation from standard input.

```bash
biblicus query --corpus corpora/example --query "primary button style preference" \\
  | biblicus context-pack build --ordering score --include-metadata --max-characters 500
```

## Reproducibility checklist

- Keep the retrieval result JSON alongside the context pack output.
- Record the policy values (`join_with`, `ordering`, `include_metadata`).
- Record any budget inputs that trimmed the context pack.

## What context pack building does

- Includes only usable text evidence.
- Excludes evidence with no text payload or whitespace-only text.

## Common pitfalls

- Building context packs from different retrieval snapshots while comparing the results.
- Comparing outputs with different `ordering` or `include_metadata` values.
- Relying on token counts without recording the tokenizer identifier.

## Token budgets

Fitting context to a token budget is a separate concern. Token counting depends on a specific tokenizer and may vary by model.

Biblicus treats token budgeting as a separate stage so it can be configured, tested, and evaluated independently from retrieval and text formatting.

In Python:

```python
from biblicus.context import (
    ContextPackPolicy,
    TokenBudget,
    fit_context_pack_to_token_budget,
)

fitted_context_pack = fit_context_pack_to_token_budget(
    context_pack,
    policy=policy,
    token_budget=TokenBudget(max_tokens=500),
)
print(fitted_context_pack.text)
```

## Ordered context blocks

Some systems assemble context from doctrine, summaries, or memory sections
before budgeting. Biblicus supports that by accepting an ordered list of
context blocks instead of a retrieval result.

In Python:

```python
from biblicus.context import (
    ContextBlock,
    ContextBlockBuildRequest,
    ContextSectionBudget,
    build_context_pack_from_blocks,
)

request = ContextBlockBuildRequest(
    blocks=[
        ContextBlock(block_id="doctrine-1", section="doctrine", text="Mission paragraph", required=True),
        ContextBlock(block_id="memory-1", section="desk_memory", text="Recent assignment summary"),
        ContextBlock(block_id="memory-2", section="desk_memory", text="Older assignment summary"),
    ],
    max_tokens=20,
    section_budgets=[
        ContextSectionBudget(section="doctrine", share=0.25),
        ContextSectionBudget(section="desk_memory", share=0.75),
    ],
)
result = build_context_pack_from_blocks(request)
print(result.text)
```

The command-line interface can build the same structure from standard input:

```bash
cat context-blocks.json | biblicus context-pack build-blocks
```

The result includes final text plus the included and dropped blocks, section
token counts, and applied budgets.
