Practical notes: Replace Your RAG Agent’s Brain With a 20B Model. It Got Smarter.

The following notes reconstruct a practical path around “Replace Your RAG Agent’s Brain With a 20B Model. It Got Smarter.”. Emphasis stays on contracts, checks, and drop-in code placeholders rather than motivational framing. When working through Overview, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

What Is Chroma Context-1?

What Is Chroma Context-1? works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

Why Traditional Agentic RAG Falls Short

Why Traditional Agentic RAG Falls Short works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

The Standard Approach

The Standard Approach works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices. The Standard Approach works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

User Query → Embedding Search → Top-K Chunks → LLM → Response

The Agentic Upgrade

For The Agentic Upgrade, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

User Query → Plan → Search → Observe → Need more info? → Search again → Generate

How Context-1 Actually Works

For How Context-1 Actually Works, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

The Four Native Tools

For The Four Native Tools, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call. For The Four Native Tools, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

Turn 1: search_corpus("merger acquisition penalties 2024")
  → Retrieves 8 chunks [Token usage: 8,432/32,768]
Turn 2: read_document("doc_14") + search_corpus("SEC filing penalties")
  → Retrieves 4 more chunks [Token usage: 18,203/32,768]
Turn 3: prune_chunks(["chunk_3", "chunk_7", "chunk_9", "chunk_11"])
  → Removes 4 irrelevant chunks [Token usage: 14,203/32,768]
Turn 4: search_corpus("specific penalty amounts regulatory action")
  → Retrieves 3 final chunks [Token usage: 21,847/32,768]
→ Returns ranked supporting documents to the reasoning model

The Self-Editing Context: Why This Is a Big Deal

When working through The Self-Editing Context: Why This Is a Big Deal, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn.

The Three-Tier Architecture: Where Context-1 Fits

When working through The Three-Tier Architecture: Where Context-1 Fits, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn.

Training: How They Built a Retrieval Specialist

When working through Training: How They Built a Retrieval Specialist, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn. When working through Training: How They Built a Retrieval Specialist, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

Phase 1: Supervised Fine-Tuning

Phase 1: Supervised Fine-Tuning works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

Phase 2: Reinforcement Learning with CISPO

Phase 2: Reinforcement Learning with CISPO works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

Reward = F1_score (recall-weighted)
       + trajectory_recall_bonus
       + final_answer_bonus (+1.0)
       - repeated_pruning_penalty (0.1 per excess)
       - turn_count_penalty

The Data Generation Pipeline: Build Your Own

The Data Generation Pipeline: Build Your Own works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices. The Data Generation Pipeline: Build Your Own works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

How the Pipeline Works

For How the Pipeline Works, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

Seed Topic → Explore Web → Extract Verifiable Facts → Generate Tasks
     ↓
Add Distractors → Chain Into Multi-Hop → Verify Answers

Verification Strategy

For Verification Strategy, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

How to Actually Use Context-1 Today

For How to Actually Use Context-1 Today, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call. For How to Actually Use Context-1 Today, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

Option 1: The API (Waitlist)

When working through Option 1: The API (Waitlist), write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn.

Option 2: Run It Yourself (With a Caveat)

When working through Option 2: Run It Yourself (With a Caveat), write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn.

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
    "chromadb/context-1",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("chromadb/context-1")

Option 3: Build Your Own Harness

When working through Option 3: Build Your Own Harness, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn. When working through Option 3: Build Your Own Harness, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

# Pseudocode for a basic Context-1 harness

tools = {
    "search_corpus": hybrid_bm25_vector_search,    # BM25 + dense vector via RRF
    "grep_corpus": regex_pattern_search,             # Regex matching, max 5 chunks
    "read_document": full_document_retrieval,        # Full doc by ID
    "prune_chunks": context_pruner                   # Remove chunks from context
}
context_budget = 32_768  # tokens
soft_threshold = 24_000
current_usage = 0
while not done:
    # Get model's next action
    response = model.generate(conversation_history)
    # Execute tool calls (model may call multiple in parallel)
    for tool_call in response.tool_calls:
        result = tools[tool_call.name](**tool_call.args)
        conversation_history.append(result)
        current_usage = count_tokens(conversation_history)
    # Enforce context budget
    if current_usage > soft_threshold:
        # Suggest pruning or concluding
        conversation_history.append(
            f"[Token usage: {current_usage}/{context_budget}] "
            "Consider pruning irrelevant chunks or concluding search."
        )

What Context-1 Can’t Do (Yet)

What Context-1 Can’t Do (Yet) works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

Why This Matters Beyond Chroma

Why This Matters Beyond Chroma works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

Over to You

Over to You works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices. Over to You works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

Let’s Keep Learning Together

For Let’s Keep Learning Together, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

Operational checklist

For Operational checklist, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state.

Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

Measure recall on a fixed question set before tuning prompts. Prompt churn rarely fixes a weak retrieval surface.

Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts.

Add a smoke test that exercises the critical path in CI with fixtures, not live paid APIs, whenever budgets allow.

Before promoting the stack, freeze versions, capture a golden transcript for the critical path, and confirm rollback steps. Shared environments need rate limits, tenancy checks, and a clear owner for secret rotation. Prefer boring reliability over clever one-off demos.

Batch note for 3ea48d84c2c9: keep provider keys out of the repo, set a per-session token ceiling, and store transcripts next to the eval fixtures so later model swaps stay comparable.