trustgraph/docs/tech-specs/agent-explainability.md
2026-03-11 14:25:28 +00:00

262 lines
9.2 KiB
Markdown

# Agent Explainability: Provenance Recording
## Overview
Add provenance recording to the React agent loop so agent sessions can be traced and debugged using the same explainability infrastructure as GraphRAG.
**Design Decisions:**
- Write to `urn:graph:retrieval` (generic explainability graph)
- Linear dependency chain for now (iteration N → wasDerivedFrom → iteration N-1)
- Tools are opaque black boxes (record input/output only)
- DAG support deferred to future iteration
## Provenance Model
```
AgentSession (urn:trustgraph:agent:{uuid})
│ tg:query = "User's question"
│ prov:startedAtTime = timestamp
│ rdf:type = tg:AgentSession
↓ prov:wasGeneratedBy
Iteration1 (urn:trustgraph:agent:{uuid}/i1)
│ tg:thought = "I need to query the knowledge base..."
│ tg:action = "knowledge-query"
│ tg:arguments = {"question": "..."}
│ tg:observation = "Result from tool..."
│ rdf:type = tg:AgentIteration
↓ prov:wasDerivedFrom
Iteration2 (urn:trustgraph:agent:{uuid}/i2)
│ ...
↓ prov:wasDerivedFrom
FinalAnswer (urn:trustgraph:agent:{uuid}/final)
│ tg:answer = "The final response..."
│ rdf:type = tg:AgentFinal
```
## Changes Required
### 1. Schema Changes
**File:** `trustgraph-base/trustgraph/schema/services/agent.py`
Add `session_id` and `collection` fields to `AgentRequest`:
```python
@dataclass
class AgentRequest:
question: str = ""
state: str = ""
group: list[str] | None = None
history: list[AgentStep] = field(default_factory=list)
user: str = ""
collection: str = "default" # NEW: Collection for provenance traces
streaming: bool = False
session_id: str = "" # NEW: For provenance tracking across iterations
```
**File:** `trustgraph-base/trustgraph/messaging/translators/agent.py`
Update translator to handle `session_id` and `collection` in both `to_pulsar()` and `from_pulsar()`.
### 2. Add Explainability Producer to Agent Service
**File:** `trustgraph-flow/trustgraph/agent/react/service.py`
Register an "explainability" producer (same pattern as GraphRAG):
```python
from ... base import ProducerSpec
from ... schema import Triples
# In __init__:
self.register_specification(
ProducerSpec(
name = "explainability",
schema = Triples,
)
)
```
### 3. Provenance Triple Generation
**Option A:** Add to existing `trustgraph-base/trustgraph/provenance/` module
**Option B:** Create agent-specific provenance helpers in the agent module
Create helper functions (similar to GraphRAG's `question_triples`, `exploration_triples`, etc.):
```python
def agent_session_triples(session_uri, query, timestamp):
"""Generate triples for agent session start."""
return [
Triple(s=session_uri, p=RDF_TYPE, o=TG_AGENT_SESSION),
Triple(s=session_uri, p=TG_QUERY, o=query),
Triple(s=session_uri, p=PROV_STARTED_AT_TIME, o=timestamp),
]
def agent_iteration_triples(iteration_uri, parent_uri, thought, action, arguments, observation):
"""Generate triples for one agent iteration."""
return [
Triple(s=iteration_uri, p=RDF_TYPE, o=TG_AGENT_ITERATION),
Triple(s=iteration_uri, p=TG_THOUGHT, o=thought),
Triple(s=iteration_uri, p=TG_ACTION, o=action),
Triple(s=iteration_uri, p=TG_ARGUMENTS, o=json.dumps(arguments)),
Triple(s=iteration_uri, p=TG_OBSERVATION, o=observation),
Triple(s=iteration_uri, p=PROV_WAS_DERIVED_FROM, o=parent_uri),
]
def agent_final_triples(final_uri, parent_uri, answer):
"""Generate triples for agent final answer."""
return [
Triple(s=final_uri, p=RDF_TYPE, o=TG_AGENT_FINAL),
Triple(s=final_uri, p=TG_ANSWER, o=answer),
Triple(s=final_uri, p=PROV_WAS_DERIVED_FROM, o=parent_uri),
]
```
### 4. Integration in service.py
**In `agent_request()` method:**
```python
async def agent_request(self, request, respond, next, flow):
# Generate or retrieve session ID
if not request.session_id:
session_id = str(uuid.uuid4())
else:
session_id = request.session_id
session_uri = f"urn:trustgraph:agent:{session_id}"
iteration_num = len(history) + 1
iteration_uri = f"{session_uri}/i{iteration_num}"
# On first iteration, emit session triples
if iteration_num == 1:
triples = agent_session_triples(session_uri, request.question, timestamp)
await flow("explainability").send(Triples(
metadata=Metadata(user=request.user, collection=..., id=session_uri),
triples=triples,
))
# ... existing react() call ...
if isinstance(act, Final):
# Emit final answer triples
final_uri = f"{session_uri}/final"
parent_uri = f"{session_uri}/i{iteration_num - 1}" if iteration_num > 1 else session_uri
triples = agent_final_triples(final_uri, parent_uri, act.final)
await flow("explainability").send(...)
else:
# Emit iteration triples
parent_uri = f"{session_uri}/i{iteration_num - 1}" if iteration_num > 1 else session_uri
triples = agent_iteration_triples(iteration_uri, parent_uri, act.thought, act.name, act.arguments, act.observation)
await flow("explainability").send(...)
# Pass session_id to next iteration
r = AgentRequest(
...,
session_id=session_id,
)
```
### 5. Predicate Definitions
**File:** `trustgraph-base/trustgraph/provenance/namespaces.py`
Add new predicates:
```python
TG_THOUGHT = TG + "thought"
TG_ACTION = TG + "action"
TG_ARGUMENTS = TG + "arguments"
TG_OBSERVATION = TG + "observation"
TG_AGENT_SESSION = TG + "AgentSession"
TG_AGENT_ITERATION = TG + "AgentIteration"
TG_AGENT_FINAL = TG + "AgentFinal"
# TG_QUERY and TG_ANSWER likely already exist
```
## Files to Modify
| File | Change |
|------|--------|
| `trustgraph-base/trustgraph/schema/services/agent.py` | Add session_id and collection to AgentRequest |
| `trustgraph-base/trustgraph/messaging/translators/agent.py` | Update translator for new fields |
| `trustgraph-base/trustgraph/provenance/namespaces.py` | Add agent predicates |
| `trustgraph-base/trustgraph/provenance/__init__.py` | Export new predicates |
| `trustgraph-flow/trustgraph/agent/react/service.py` | Add explainability producer + recording logic |
| `trustgraph-cli/trustgraph/cli/show_explain_trace.py` | Handle agent trace types |
| `trustgraph-cli/trustgraph/cli/list_explain_traces.py` | List agent sessions alongside GraphRAG |
## Files to Potentially Create
| File | Purpose |
|------|---------|
| `trustgraph-base/trustgraph/provenance/agent.py` | Agent-specific triple generators (optional, could inline in service.py) |
## CLI Updates Detail
**`list_explain_traces.py`:**
- Currently queries for `tg:query` predicate to find questions
- Agent sessions also use `tg:query`, so should work automatically
- May want to add a type indicator column (GraphRAG vs Agent)
**`show_explain_trace.py`:**
- Currently expects: question → exploration → focus → synthesis chain
- Agent traces have: session → iteration(s) → final chain
- Detection: check `rdf:type` of the root entity
- `tg:AgentSession` → agent trace rendering
- Otherwise → GraphRAG trace rendering (existing)
- Agent rendering shows:
- Session info (question, time)
- Each iteration: thought, action, args, observation
- Final answer
## Design Decisions
1. **Collection handling:** Add `collection` field to AgentRequest. Agent receives a collection parameter for provenance traces. Tools can access other collections per their config, but decision traces stay in the invoked collection.
2. **CLI tool updates:** Update existing `tg-show-explain-trace` to detect and handle both GraphRAG and agent trace types.
## Implementation Order
0. **Tech Spec:** Write this plan to `docs/tech-specs/agent-explainability.md`
1. **Schema:** Add `session_id` and `collection` to AgentRequest + translator
2. **Predicates:** Add agent predicates to `namespaces.py`
3. **Service:** Add explainability producer to agent service
4. **Provenance:** Create/add agent triple generators
5. **Integration:** Wire up provenance recording in `agent_request()`
6. **CLI:** Update `show_explain_trace.py` to detect and render agent traces
7. **Test:** Run agent query and verify trace is recorded and viewable
## Backwards Compatibility
- `session_id` defaults to `""` - old requests work, just won't have provenance
- `collection` defaults to `"default"` - reasonable fallback
- CLI gracefully handles both trace types
## Verification
```bash
# Run an agent query
tg-invoke-agent -q "What is the capital of France?"
# Check triples were written
tg-query-graph -p "https://trustgraph.ai/ns/query" -g "urn:graph:retrieval"
# List traces (should show agent sessions)
tg-list-explain-traces -U trustgraph -C default
# Show trace (may need CLI updates for agent entity types)
tg-show-explain-trace "urn:trustgraph:agent:xxx"
```
## Future Work (Not This PR)
- DAG dependencies (when iteration N uses results from multiple prior iterations)
- Tool-specific provenance linking (KnowledgeQuery → its GraphRAG trace)
- CLI tool enhancements to better display agent traces
- Streaming provenance emission (emit as we go, not batch at end)