fic-agent

A no-finetune, retrieval-augmented character chatbot for fiction.

Goals

• Factual consistency with the source text
• Character-true language style and emotional profile
• Worldview consistency (settings, rules, relationships)

Repository layout

• configs/ Runtime configs (API keys, model names, flags)
• data/raw/ Raw novels and metadata
• data/processed/ Chunked text, extracted dialogues, persona notes
• data/indexes/ Vector indexes and graph artifacts
• data/eval/ Evaluation inputs and references
• outputs/ Run outputs and logs
• scripts/ CLI helpers and pipelines
• src/fic_agent/ Core library

Core pipeline (high level)

1. Ingest raw text
2. Extract character dialogue + context
3. Build persona profile (style, affect, worldview)
4. Build fact and worldview indexes
5. Retrieve relevant evidence + persona
6. Generate response with persona-conditioned prompt
7. Evaluate (facts, style, worldview)

End-to-End Run (Dataset -> Question -> Answer)

1. Activate environment and install dependencies.

conda activate fic_agent
cd /home/tleautomat/Project/fic-agent
python -m pip install -r requirements.txt

2. Set API keys (OpenRouter key can be reused for LLM and embeddings).

export OPENROUTER_API_KEY="your_key"
export OPENAI_API_KEY="$OPENROUTER_API_KEY"
export EMBEDDING_API_KEY="$OPENROUTER_API_KEY"

3. Put your novel text in data/raw/ (example already included).

4. Run preprocessing + memory building + index building.

python scripts/run_pipeline.py \
  --input "data/raw/The Strange Case of Dr. Jekyll and Mr. Hyde.txt" \
  --book-id jekyll_hyde \
  --characters "Jekyll,Hyde,Utterson,Poole" \
  --all-characters \
  --worldview-llm \
  --build-index

5. Ask a user question with the meta-cognitive QA loop.

python scripts/run_meta_qa.py \
  --query "As Utterson, should you trust Hyde after the trampling incident?" \
  --character Utterson \
  --max-iter 3 \
  --style-correct \
  --dump-trace \
  --save-json outputs/qa/utterson_q1.json

6. (Optional) Evaluate one generated answer with the new comprehensive LLM judge.

PYTHONPATH=src python scripts/eval_answer.py \
  --result-json outputs/qa/utterson_q1.json \
  --character Utterson \
  --processed-dir data/processed \
  --mode llm \
  --rounds 3 \
  --temperature 0.2 \
  --top-n 6 \
  --save-json outputs/qa/utterson_q1_eval.json

Evaluation (LLM Comprehensive Judge)

The project now uses a comprehensive LLM evaluator (v2) that combines canon grounding, character realization, world logic, and answer usefulness.

Core modules (each sub-score uses a 1-5 scale):

• canon_grounding: evidence_fit, fact_consistency, timeline_consistency
• character_realization: linguistic_style, tone, vocabulary, emotional_expression, motivation_alignment
• world_logic: rule_consistency, setting_relation_consistency
• response_usefulness: question_fit, clarity

Outputs include:

• scorecard.overall_100 (weighted by 40/35/15/10)
• same_character (Yes/No)
• confidence_100 (0-100)
• issues.critical/major/minor
• penalties and per-round breakdown
• Backward-compatible fields: scores.facts/persona/worldview/overall and confidence (0-1)

Run:

PYTHONPATH=src python scripts/eval_answer.py \
  --result-json outputs/qa/utterson_q1.json \
  --character Utterson \
  --processed-dir data/processed \
  --mode llm \
  --rounds 3

Note:

• If you already installed with pip install -e ., you can omit PYTHONPATH=src.

Main Outputs

• data/processed/chunks.jsonl: document chunks
• data/processed/dialogues.jsonl: extracted dialogues with speaker guesses
• data/processed/persona_*.json: persona profiles per character
• data/processed/worldview_notes.jsonl: worldview memory notes
• data/indexes/*.faiss and data/indexes/*.jsonl: vector indexes + metadata
• outputs/qa/*.json: final answer, loop trace, merged evidence
• outputs/qa/*_eval.json: comprehensive LLM evaluation reports (including same_character, confidence_100, and penalties)

Dataset Scene Experiments (105_Persuasion)

The new batch script runs the following protocol:

• Iterate all scenes with is_two_person_dialogue_scene=true in ficset/105_Persuasion
• Use all prior scenes as RAG knowledge
• Use the first speaker's utterance as query input
• Ask the model to answer as the other character (reusing existing retrieval and meta-loop logic)

Dry-run (no model API calls, just case construction):

cd /home/tleautomat/Project/fic-agent
PYTHONPATH=src python scripts/run_scene_experiments.py \
  --book 105_Persuasion \
  --dry-run \
  --run-name exp105_dryrun

Full run (all true scenes):

cd /home/tleautomat/Project/fic-agent
bash scripts/run_105_persuasion_experiments.sh \
  --run-name exp105_full \
  --style-correct

Optional: run per-case evaluation in the same batch

bash scripts/run_105_persuasion_experiments.sh \
  --run-name exp105_eval \
  --eval-mode llm \
  --eval-rounds 3

Outputs:

• outputs/experiments/<run_name>/summary.json: global summary and per-case status
• outputs/experiments/<run_name>/cases.jsonl: per-case records (line-delimited)
• outputs/experiments/<run_name>/cases/<case_key>/qa_full.json: QA result for one case
• outputs/experiments/<run_name>/cases/<case_key>/eval_full.json: evaluation result (if enabled)