fabra-ai 2.2.0

Context infrastructure for AI applications

Fabra

Context Records for LLM incidents.

Capture what the model saw as a durable artifact you can replay, verify, and diff.
record -> replay -> diff

---

When An LLM Incident Happens

• Support shares a screenshot.
• Engineering tries to reconstruct prompts, retrieval, features, and runtime state.
• Nobody can answer, precisely and repeatably: what did it see, what changed, why did it answer that way?

Fabra's unit of evidence is a context_id (UUID or ctx_<UUID>). Paste it into the ticket, then:

fabra context show <context_id>
fabra context verify <context_id>
fabra context diff <context_id_A> <context_id_B>

---

Quickstart (No Keys, No Docker)

pip install fabra-ai
fabra demo

fabra demo starts a local server, makes a test request, and prints a working context_id plus the next commands to run.

By default, Context Records are stored durably in DuckDB at ~/.fabra/fabra.duckdb (override with FABRA_DUCKDB_PATH).

Requirements

• Python >= 3.9
• pip (or uv, optional)
• curl (used in quickstart verification commands)
• Optional:
  • Node.js (only for fabra ui)
  • Docker (only for Docker-based examples / local production stack)

Help and diagnostics:

fabra --help
fabra context --help
fabra doctor

Already using LangChain or calling OpenAI directly? Add receipts without changing your app architecture:

• docs/exporters-and-adapters.md (LangChain-style callbacks, OpenAI wrapper, logs/OTEL)

---

What A Context Record Is

A Context Record is a durable artifact for an AI request: assembled content plus lineage (features, retrieval, freshness decisions), and optional integrity metadata.

Create a context_id locally using the shipped example:

fabra serve examples/demo_context.py

In another terminal:

curl -fsS -X POST http://127.0.0.1:8000/v1/context/chat_context \
  -H "Content-Type: application/json" \
  -d '{"user_id":"u1","query":"test"}'

Core workflow (from a context_id):

fabra context show <context_id>
fabra context verify <context_id>
fabra context diff <context_id_A> <context_id_B>
fabra context export <context_id> --bundle -o incident_bundle.zip

incident_bundle.zip includes the exported record (<context_id>.json) plus manifest.json with stored vs computed hashes.

Drop-In Receipts (No Keys)

If you already build prompts but don’t use @context yet, emit a verifiable receipt:

from fabra.receipts import ReceiptRecorder
from fabra.exporters.logging import emit_context_id_json

prompt = "system: ...\nuser: ..."
recorder = ReceiptRecorder()
receipt = recorder.record_sync(
    context_function="my_llm_call",
    content=prompt,
    inputs={"model": "gpt-4.1-mini"},
)
emit_context_id_json(receipt.context_id, source="my-service", ticket="INC-1234")

---

Two Entry Points

ML Engineers (Features) Define features in Python and serve them over HTTP. from fabra import FeatureStore, entity, feature from datetime import timedelta store = FeatureStore() @entity(store) class User: user_id: str @feature(entity=User, refresh=timedelta(hours=1)) def purchase_count(user_id: str) -> int: # Replace with a real DB call return 47 fabra serve features.py # or: fabra serve examples/demo_features.py curl localhost:8000/features/purchase_count?entity_id=u123 # {"value": 47, "freshness_ms": 0, "served_from": "online"}

AI Engineers (Context) Assemble prompts with features + retrieval, with lineage and budgets. from fabra import FeatureStore from fabra.context import context, ContextItem store = FeatureStore() @context(store, max_tokens=4000, freshness_sla="5m") async def build_prompt(user_id: str, query: str) -> list[ContextItem]: # Replace with your feature calls and retriever / RAG call tier = "premium" docs = "..." return [ ContextItem(content=f"User tier: {tier}", priority=1), ContextItem(content=docs, priority=2), ] ctx = await build_prompt("user_123", "question") print(ctx.id) print(ctx.lineage)

---

CLI (Most Used)

fabra demo                          # start demo + print a context_id
fabra context show <id>             # inspect a record (or best-effort legacy view)
fabra context verify <id>           # verify CRS-001 hashes (fails if unavailable)
fabra context diff <a> <b>          # compare two contexts
fabra context export <id>           # export json/yaml
fabra context export <id> --bundle  # zip bundle for incident/audit
fabra doctor                        # local diagnostics
fabra serve <file.py>               # run server for examples/your code
fabra deploy fly|cloudrun|ecs        # generate deployment config
fabra ui <file.py>                  # local UI (requires Node.js)

Note: fabra ui requires Node.js. If you’re running from source, install UI deps with cd src/fabra/ui-next && npm install.

---

What Fabra Is / Isn’t

Fabra is infrastructure for inference-time evidence (Context Records) and serving (features/context).

Fabra is not:

• an agent framework
• a monitoring dashboard
• a no-code builder

See docs/quickstart.md, docs/context-record-spec.md, and docs/comparisons.md for details.

---

Try in Browser · Quickstart · Docs